Spectra of Graph by Haemers

Spectra of graphs
Andries E. Brouwer Willem H. Haemers

2
Contents
1 Graph spectrum 11
1.1 Matrices associated to a graph . . . . . . . . . . . . . . . . . . . 11
1.2 The spectrum of a graph . . . . . . . . . . . . . . . . . . . . . . . 11
1.2.1 Characteristic polynomial . . . . . . . . . . . . . . . . . . 13
1.3 The spectrum of an undirected graph . . . . . . . . . . . . . . . . 13
1.3.1 Regular graphs . . . . . . . . . . . . . . . . . . . . . . . . 13
1.3.2 Complements . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.3.3 Walks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.3.4 Diameter . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.3.5 Spanning trees . . . . . . . . . . . . . . . . . . . . . . . . 14
1.3.6 Bipartite graphs . . . . . . . . . . . . . . . . . . . . . . . 15
1.3.7 Connectedness . . . . . . . . . . . . . . . . . . . . . . . . 16
1.4 Spectrum of some graphs . . . . . . . . . . . . . . . . . . . . . . 17
1.4.1 The complete graph . . . . . . . . . . . . . . . . . . . . . 17
1.4.2 The complete bipartite graph . . . . . . . . . . . . . . . . 17
1.4.3 The cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.4.4 The path . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.4.5 Line graphs . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.4.6 Cartesian products . . . . . . . . . . . . . . . . . . . . . . 18
1.4.7 Kronecker products and bipartite double . . . . . . . . . . 19
1.4.8 Cayley graphs . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.5 Decompositions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.5.1 Decomposing K10 into Petersen graphs . . . . . . . . . . 20
1.5.2 Decomposing Kn into complete bipartite graphs . . . . . 20
1.6 Automorphisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.7 Algebraic connectivity . . . . . . . . . . . . . . . . . . . . . . . . 21
1.8 Cospectral graphs . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.8.1 The 4-cube . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.8.2 Seidel switching . . . . . . . . . . . . . . . . . . . . . . . . 23
1.8.3 Godsil-McKay switching . . . . . . . . . . . . . . . . . . . 23
1.8.4 Reconstruction . . . . . . . . . . . . . . . . . . . . . . . . 24
1.9 Very small graphs . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.10 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2 Linear algebra 27
2.1 Simultaneous diagonalization . . . . . . . . . . . . . . . . . . . . 27
2.2 Perron-Frobenius Theory . . . . . . . . . . . . . . . . . . . . . . 27
2.3 Equitable partitions . . . . . . . . . . . . . . . . . . . . . . . . . 30
3
4 CONTENTS
2.3.1 Equitable and almost equitable partitions of graphs . . . 30

2.4 The Rayleigh quotient . . . . . . . . . . . . . . . . . . . . . . . . 31
2.5 Interlacing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.6 Schur’s inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.7 Schur complements . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.8 The Courant-Weyl inequalities . . . . . . . . . . . . . . . . . . . 34
2.9 Gram matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.10 Diagonally dominant matrices . . . . . . . . . . . . . . . . . . . . 36
2.10.1 Geršgorin circles . . . . . . . . . . . . . . . . . . . . . . . 36
2.11 Projections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.12 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3 Graph spectrum (II) 39

3.1 The largest eigenvalue . . . . . . . . . . . . . . . . . . . . . . . . 39
3.1.1 Graphs with largest eigenvalue at most 2 . . . . . . . . . 40
3.1.2 Subdividing an edge . . . . . . . . . . . . . . . . . . . . . 41
3.1.3 The Kelmans operation . . . . . . . . . . . . . . . . . . . 42
3.2 Interlacing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.3 Regular graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.4 Bipartite graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.5 Classification of integral cubic graphs . . . . . . . . . . . . . . . . 44
3.6 The largest Laplace eigenvalue . . . . . . . . . . . . . . . . . . . 47
3.7 Laplace eigenvalues and degrees . . . . . . . . . . . . . . . . . . . 48
3.8 The Grone-Merris Conjecture . . . . . . . . . . . . . . . . . . . . 50
3.8.1 Threshold graphs . . . . . . . . . . . . . . . . . . . . . . . 50
3.8.2 Proof of the Grone-Merris Conjecture . . . . . . . . . . . 50
3.9 The Laplacian for hypergraphs . . . . . . . . . . . . . . . . . . . 53
3.10 Applications of eigenvectors . . . . . . . . . . . . . . . . . . . . . 55
3.10.1 Ranking . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.10.2 Google Page rank . . . . . . . . . . . . . . . . . . . . . . . 55
3.10.3 Cutting . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.10.4 Graph drawing . . . . . . . . . . . . . . . . . . . . . . . . 56
3.10.5 Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.10.6 Searching an eigenspace . . . . . . . . . . . . . . . . . . . 58
3.11 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4 Substructures 61
4.1 Cliques and cocliques . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.1.1 Using weighted adjacency matrices . . . . . . . . . . . . . 62
4.2 Chromatic number . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.2.1 Using weighted adjacency matrices . . . . . . . . . . . . . 64
4.2.2 Rank and chromatic number . . . . . . . . . . . . . . . . 64
4.3 Shannon capacity . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.4 Stars and star complements . . . . . . . . . . . . . . . . . . . . . 65
4.5 The second largest eigenvalue . . . . . . . . . . . . . . . . . . . . 66
4.5.1 Bounds for the second largest eigenvalue . . . . . . . . . . 66
4.5.2 Large regular subgraphs are connected . . . . . . . . . . . 67
4.5.3 Randomness . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.5.4 Expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.5.5 Toughness and Hamiltonicity . . . . . . . . . . . . . . . . 69
CONTENTS 5
4.5.6 The Petersen graph is not Hamiltonian . . . . . . . . . . 70

4.5.7 Diameter bound . . . . . . . . . . . . . . . . . . . . . . . 70
4.6 Separation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.6.1 Bandwidth . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.6.2 Perfect matchings . . . . . . . . . . . . . . . . . . . . . . 72
4.7 Block Designs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.8 Polarities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.9 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5 Trees 79
5.1 Characteristic polynomials of trees . . . . . . . . . . . . . . . . . 79
5.2 Eigenvectors and multiplicities . . . . . . . . . . . . . . . . . . . 81
5.3 Sign patterns of eigenvectors of graphs . . . . . . . . . . . . . . . 82
5.4 Sign patterns of eigenvectors of trees . . . . . . . . . . . . . . . . 83
5.5 The spectral center of a tree . . . . . . . . . . . . . . . . . . . . . 84
5.6 Integral trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6 Groups and graphs 87

6.1 Γ(G, H, S) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.2 Spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.3 Covers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
6.4 Cayley sum graphs . . . . . . . . . . . . . . . . . . . . . . . . . . 90
6.4.1 (3,6)-fullerenes . . . . . . . . . . . . . . . . . . . . . . . . 90
6.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
7 Euclidean representations 93
7.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
7.2 Euclidean representation . . . . . . . . . . . . . . . . . . . . . . . 93
7.3 Root lattices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
7.4 Cameron-Goethals-Seidel-Shult . . . . . . . . . . . . . . . . . . . 99
7.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
8 Strongly regular graphs 101

8.1 Strongly regular graphs . . . . . . . . . . . . . . . . . . . . . . . 101
8.1.1 Simple examples . . . . . . . . . . . . . . . . . . . . . . . 101
8.1.2 The Paley graphs . . . . . . . . . . . . . . . . . . . . . . . 102
8.1.3 Adjacency matrix . . . . . . . . . . . . . . . . . . . . . . 103
8.1.4 Imprimitive graphs . . . . . . . . . . . . . . . . . . . . . . 103
8.1.5 Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . 103
8.1.6 The half case and cyclic strongly regular graphs . . . . . 104
8.1.7 Strongly regular graphs without triangles . . . . . . . . . 104
8.1.8 Further parameter restrictions . . . . . . . . . . . . . . . 105
8.1.9 Strongly regular graphs from permutation groups . . . . . 106
8.1.10 Strongly regular graphs from quasisymmetric designs . . . 107
8.1.11 Symmetric 2-designs from strongly regular graphs . . . . 107
8.1.12 Latin square graphs . . . . . . . . . . . . . . . . . . . . . 107
8.1.13 Partial Geometries . . . . . . . . . . . . . . . . . . . . . . 108
8.2 Strongly regular graphs with eigenvalue −2 . . . . . . . . . . . . 109
8.3 Connectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
6 CONTENTS
8.4 Cocliques and colorings . . . . . . . . . . . . . . . . . . . . . . . 112

8.5 Automorphisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
8.6 Generalized quadrangles . . . . . . . . . . . . . . . . . . . . . . . 114
8.6.1 Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . 114
8.6.2 Constructions of generalized quadrangles . . . . . . . . . 115
8.6.3 Strongly regular graphs from generalized quadrangles . . 116
8.6.4 Generalized quadrangles with lines of size 3 . . . . . . . . 116
8.7 The (81,20,1,6) strongly regular graph . . . . . . . . . . . . . . . 117
8.7.1 Descriptions . . . . . . . . . . . . . . . . . . . . . . . . . . 117
8.7.2 Uniqueness . . . . . . . . . . . . . . . . . . . . . . . . . . 118
8.7.3 Independence and chromatic numbers . . . . . . . . . . . 119
8.8 Strongly regular graphs and 2-weight codes . . . . . . . . . . . . 120
8.8.1 Codes, graphs and projective sets . . . . . . . . . . . . . . 120
8.8.2 The correspondence between linear codes and subsets of
a projective space . . . . . . . . . . . . . . . . . . . . . . 120
8.8.3 The correspondence between projective two-weight codes,
subsets of a projective space with two intersection num-
bers, and affine strongly regular graphs . . . . . . . . . . 121
8.8.4 Duality for affine strongly regular graphs . . . . . . . . . 123
8.8.5 Cyclotomy . . . . . . . . . . . . . . . . . . . . . . . . . . 124
8.9 Table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
8.10 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
9 Regular two-graphs 133

9.1 Strong graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
9.2 Two-graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
9.3 Regular two-graphs . . . . . . . . . . . . . . . . . . . . . . . . . . 135
9.3.1 Related strongly regular graphs . . . . . . . . . . . . . . . 137
9.3.2 The regular two-graph on 276 points . . . . . . . . . . . . 138
9.3.3 Coherent subsets . . . . . . . . . . . . . . . . . . . . . . . 138
9.3.4 Completely regular two-graphs . . . . . . . . . . . . . . . 138
9.4 Conference matrices . . . . . . . . . . . . . . . . . . . . . . . . . 139
9.5 Hadamard matrices . . . . . . . . . . . . . . . . . . . . . . . . . . 140
9.6 Equiangular lines . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
9.6.1 Equiangular lines in Rd and two-graphs . . . . . . . . . . 142
9.6.2 Bounds on equiangular sets of lines in Rd or Cd . . . . . . 143
9.6.3 Bounds on sets of lines with few angles and sets of vectors
with few distances . . . . . . . . . . . . . . . . . . . . . . 144
10 Association schemes 147

10.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
10.2 The Bose-Mesner algebra . . . . . . . . . . . . . . . . . . . . . . 148
10.3 The Linear Programming Bound . . . . . . . . . . . . . . . . . . 150
10.4 The Krein parameters . . . . . . . . . . . . . . . . . . . . . . . . 151
10.5 Automorphisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
10.5.1 The Moore graph on 3250 vertices . . . . . . . . . . . . . 153
10.6 P - and Q-polynomial association schemes . . . . . . . . . . . . . 153
10.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
CONTENTS 7
11 Distance regular graphs 157

11.1 Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
11.2 Spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
11.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
11.3.1 Hamming graphs . . . . . . . . . . . . . . . . . . . . . . . 158
11.3.2 Johnson graphs . . . . . . . . . . . . . . . . . . . . . . . . 159
11.3.3 Grassmann graphs . . . . . . . . . . . . . . . . . . . . . . 159
11.3.4 Van Dam-Koolen graphs . . . . . . . . . . . . . . . . . . . 159
11.4 Bannai-Ito conjecture . . . . . . . . . . . . . . . . . . . . . . . . 160
11.5 Connectedness . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
11.6 Degree of eigenvalues . . . . . . . . . . . . . . . . . . . . . . . . . 160
11.7 Moore graphs and generalized polygons . . . . . . . . . . . . . . 161
11.8 Primitivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
11.9 Euclidean representations . . . . . . . . . . . . . . . . . . . . . . 162
11.10Extremality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
11.11Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
12 p-ranks 165
12.1 Reduction mod p . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
12.2 The minimal polynomial . . . . . . . . . . . . . . . . . . . . . . . 166
12.3 Bounds for the p-rank . . . . . . . . . . . . . . . . . . . . . . . . 166
12.4 Interesting primes p . . . . . . . . . . . . . . . . . . . . . . . . . 167
12.5 Adding a multiple of J . . . . . . . . . . . . . . . . . . . . . . . . 167
12.6 Paley graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
12.7 Strongly regular graphs . . . . . . . . . . . . . . . . . . . . . . . 169
12.8 Smith Normal Form . . . . . . . . . . . . . . . . . . . . . . . . . 171
12.8.1 Smith Normal Form and spectrum . . . . . . . . . . . . . 172
13 Spectral characterizations 175

13.1 Generalized adjacency matrices . . . . . . . . . . . . . . . . . . . 175
13.2 Constructing cospectral graphs . . . . . . . . . . . . . . . . . . . 176
13.2.1 Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
13.2.2 Partial linear spaces . . . . . . . . . . . . . . . . . . . . . 177
13.2.3 GM switching . . . . . . . . . . . . . . . . . . . . . . . . . 178
13.2.4 Sunada’s method . . . . . . . . . . . . . . . . . . . . . . . 179
13.3 Enumeration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
13.3.1 Lower bounds . . . . . . . . . . . . . . . . . . . . . . . . . 179
13.3.2 Computer results . . . . . . . . . . . . . . . . . . . . . . . 180
13.4 DS graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
13.4.1 Spectrum and structure . . . . . . . . . . . . . . . . . . . 181
13.4.2 Some DS graphs . . . . . . . . . . . . . . . . . . . . . . . 183
13.4.3 Line graphs . . . . . . . . . . . . . . . . . . . . . . . . . . 185
13.5 Distance-regular graphs . . . . . . . . . . . . . . . . . . . . . . . 187
13.5.1 Strongly regular DS graphs . . . . . . . . . . . . . . . . . 187
13.5.2 Distance-regularity from the spectrum . . . . . . . . . . . 188
13.5.3 Distance-regular DS graphs . . . . . . . . . . . . . . . . . 189
13.6 The method of Wang & Xu . . . . . . . . . . . . . . . . . . . . . 192
13.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
8 CONTENTS
14 Graphs with few eigenvalues 195

14.1 Regular graphs with four eigenvalues . . . . . . . . . . . . . . . . 195
14.2 Three Laplacian eigenvalues . . . . . . . . . . . . . . . . . . . . . 196
14.3 Other matrices with at most three eigenvalues . . . . . . . . . . . 198
14.3.1 Few Seidel eigenvalues . . . . . . . . . . . . . . . . . . . . 198
14.3.2 Three adjacency eigenvalues . . . . . . . . . . . . . . . . . 199
14.3.3 Three signless Laplacian eigenvalues . . . . . . . . . . . . 200
14.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
Subject Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
CONTENTS 9
Algebraic graph theory is the branch of mathematics that studies graphs by

using algebraic properties of associated matrices. More in particular, spectral
graph theory studies the relation between graph properties and the spectrum of
the adjacency matrix or Laplace matrix. And the theory of association schemes
and coherent configurations studies the algebra generated by associated matri-
ces.
10 CONTENTS
Chapter 1
Graph spectrum
This chapter presents some simple results on graph spectra. We assume the
reader to be familiar with elementary linear algebra and graph theory. Through-
out J will denote the all-1 matrix, and 1 is the all-1 vector.
1.1 Matrices associated to a graph

Let Γ be a graph without multiple edges. The adjacency matrix of Γ is the
0-1 matrix A indexed by the vertex set V Γ of Γ, where Axy = 1 when there
is an edge from x to y in Γ and Axy = 0 otherwise. Occasionally we consider
multigraphs (possibly with loops) in which case Axy equals the number of edges
from x to y.
Let Γ be an undirected graph without loops. The (vertex-edge) incidence
matrix of Γ is the 0-1 matrix M , with rows indexed by the vertices and columns
indexed by the edges, where Mxe = 1 when vertex x is an endpoint of edge e.
Let Γ be a directed graph without loops. The directed incidence matrix of
Γ is the 0-1 matrix N , with rows indexed by the vertices and columns by the
edges, where Nxe = −1, 1, 0 when x is the head of e, the tail of e, or not on e,
respectively.
Let Γ be an undirected graph without loops. The Laplace matrix of Γ is the
matrix L indexed by the vertex set of Γ, with zero row sums, where Lxy = −Axy
for x 6= y. If D is the diagonal matrix, indexed by the vertex set of Γ such that
Dxx is the degree (valency) of x, then L = D − A. The matrix Q = D + A is
called the signless Laplace matrix of Γ.
An important property of the Laplace matrix L and the signless Laplace
matrix Q is that they are positive semidefinite. Indeed, one has Q = M M ⊤
and L = N N ⊤ if M is the incidence matrix of Γ and N the directed incidence
matrix of the directed graph obtained by orienting the edgesPof Γ in an arbitrary
way. It follows that for any vector u one has u⊤ Lu = 2
xy (ux − uy ) and
⊤
P 2
u Qu = xy (ux + uy ) , where the sum is over the edges of Γ.
1.2 The spectrum of a graph

The (ordinary) spectrum of a finite graph Γ is by definition the spectrum of the
adjacency matrix A, that is, its set of eigenvalues together with their multiplic-
11
12 CHAPTER 1. GRAPH SPECTRUM
ities. The Laplace spectrum of a finite undirected graph without loops is the
spectrum of the Laplace matrix L.
The rows and columns of a matrix of order n are numbered from 1 to n,
while A is indexed by the vertices of Γ, so that writing down A requires one to
assign some numbering to the vertices. However, the spectrum of the matrix
obtained does not depend on the numbering chosen. It is the spectrum of the
linear transformation A on the vector space K X of maps from X into K, where
X is the vertex set, and K is some field such as R or C.
The characteristic polynomial of Γ is that of A, that is, the polynomial pA
defined by pA (θ) = det(θI − A).
Example Let Γ be the path P3 with three vertices and two edges. Assigning
some arbitrary order to the three vertices of Γ, we find that the adjacency matrix
A becomes one of
     
0 1 1 0 1 0 0 0 1
 1 0 0  or  1 0 1  or  0 0 1  .
1 0 0 0 1 0 1 1 0
√ √
The characteristic polynomial is pA (θ) = θ3 − 2θ. The spectrum is 2, 0, − 2.
The eigenvectors are:
√ √ √ √
2 2 2 1 0 −1 2 −2 2
c c c c c c c c c .
(Here, for an eigenvector

P u, we write ux as a label at the vertex x; one has
Au = θu if and only if y←x uy = θux for all x.) The Laplace matrix L is one
of      
2 −1 −1 1 −1 0 1 0 −1
 −1 1 0  or  −1 2 −1  or  0 1 −1  .
−1 0 1 0 −1 1 −1 −1 2
Its eigenvalues are 0, 1 and 3. The Laplace eigenvectors are:
1 1 1 1 0 −1 1 −2 1
c c c c c c c c c .
P
(One has Lu = θu if and only if y∼x uy = (dx − θ)ux for all x, where dx is
the degree of the vertex x.)
Example Let Γ be the directed triangle with adjacency matrix
 
0 1 0
A =  0 0 1 .
1 0 0
Then A has characteristic polynomial pA (θ) = θ3 − 1 and spectrum 1, ω, ω 2 ,

where ω is a primitive cube root of unity.
Example Let Γ be the directed graph with two vertices and a single directed
0 1
edge. Then A = with pA (θ) = θ2 . So A has the eigenvalue 0 with
0 0
geometric multiplicity (that is, the dimension of the corresponding eigenspace)
equal to 1 and algebraic multiplicity (that is, its multiplicity as a root of the
polynomial pA ) equal to 2.
1.3. THE SPECTRUM OF AN UNDIRECTED GRAPH 13
1.2.1 Characteristic polynomial

Let Γ be a directed graph on n vertices. For any directed subgraph C of Γ
that is a union of directed cycles, let c(C) be its number of cycles. P Then the
characteristicPpolynomial pA (t) = det(tI −A) of Γ can be expanded as ci tn−i ,
where ci = C (−1)c(C) , with C running over all regular directed subgraphs
with in- and outdegree 1 on i vertices.
(Indeed,
P this is just a reformulation of the definition of the determinant as
det M = σ sgn(σ)M1σ(1) ...Mnσ(n) . Note that when the permutation σ with
n − i fixed points is written as a product of non-identity cycles, its sign is (−1)e
where e is the number of even cycles in this product. Since the number of odd
non-identity cycles is congruent to i (mod 2), we have sgn(σ) = (−1)i+c(σ) .)
For example, the directed triangle has c0 = 1, c3 = −1. Directed edges that
do not occur in directed cycles do not influence the (ordinary) spectrum.
The same description of pA (t) holds for undirected graphs (with each edge
viewed as a pair of opposite directed edges).
d
P
Since dt det(tI − A) = x det(tI − Ax ) where Ax is the submatrix of A
obtained by deleting row and column x, it follows that p′A (t) is the sum of the
characteristic polynomials of all single-vertex-deleted subgraphs of Γ.
1.3 The spectrum of an undirected graph

Suppose Γ is undirected and simple with n vertices. Since A is real and sym-
metric, all its eigenvalues are real. Also, for each eigenvalue θ, its algebraic
multiplicity coincides with its geometric multiplicity, so that we may omit the
adjective and just speak about ‘multiplicity’. Conjugate algebraic integers have
the same multiplicity. Since A has zero diagonal, its trace tr A, and hence the
sum of the eigenvalues is zero.
Similarly, L is real and symmetric, so that the Laplace spectrum is real.
Moreover, L is positive semidefinite and singular, so we may denote the eigen-
values by µ1 , . . . , µn , where 0 = µ1 ≤ µ2 ≤ . . . ≤ µn . The sum of these
eigenvalues is tr L, which is twice the number of edges of Γ.
Finally, also Q has real spectrum and nonnegative eigenvalues (but is not
necessarily singular). We have tr Q = tr L.
1.3.1 Regular graphs

A graph Γ is called regular of degree (or valency) k, when every vertex has
precisely k neighbors. So, Γ is regular of degree k precisely when its adjacency
matrix A has row sums k, i.e., when A1 = k1 (or AJ = kJ).
If Γ is regular of degree k, then for every eigenvalue θ we have |θ| ≤ k. (One
way to see this, is by observing that if |t| > k then the matrix tI − A is strictly
diagonally dominant, and hence nonsingular, so that t is not an eigenvalue of
A.)
If Γ is regular of degree k, then L = kI − A. It follows that if Γ has ordinary
eigenvalues k = θ1 ≥ . . . ≥ θn and Laplace eigenvalues 0 = µ1 ≤ µ2 ≤ . . . ≤
µn , then θi = k − µi for i = 1, . . . , n. The eigenvalues of Q = kI + A are
2k, k + θ2 , . . . , k + θn .
1.3.2 Complements
The complement Γ of Γ is the graph with the same vertex set as Γ, where two
distinct vertices are adjacent whenever they are nonadjacent in Γ. So, if Γ has
adjacency matrix A, then Γ has adjacency matrix A = J − I − A and Laplace
matrix L = nI − J − L.
Because eigenvectors of L are also eigenvectors of J, the eigenvalues of L are
0, n − µn , . . . , n − µ2 . (In particular, µn ≤ n.)
If Γ is regular we have a similar result for the ordinary eigenvalues: if Γ is k-
regular with eigenvalues θ1 ≥ . . . ≥ θn , then the eigenvalues of the complement
are n − k − 1, −1 − θn , . . . , −1 − θ2 .
1.3.3 Walks
From the spectrum one can read off the number of closed walks of a given length.
Proposition 1.3.1 Let h be a nonnegative integer. Then (Ah )xy is the number
of walks of length h from x to y. In particular, (A2 )xx is the degree of the vertex
x, and tr A2 equals twice the number of edges of Γ; similarly, tr A3 is six times
the number of triangles in Γ.
1.3.4 Diameter
We saw that all eigenvalues of a single directed edge are zero. For undirected
graphs this does not happen.
Proposition 1.3.2 Let Γ be an undirected graph. All its eigenvalues are zero
if and only if Γ has no edges. The same holds for the Laplace eigenvalues and
the signless Laplace eigenvalues.
More generally, we find a lower bound for the diameter:
Proposition 1.3.3 Let Γ be a connected graph with diameter d. Then Γ has at

least d + 1 distinct eigenvalues, at least d + 1 distinct Laplace eigenvalues, and
at least d + 1 distinct signless Laplace eigenvalues.
Proof. Let M be any nonnegative symmetric matrix with rows and columns
indexed by V Γ and such that for distinct vertices x, y we have Mxy > 0 if and
only if x ∼ y. Let the distinct eigenvalues of M be θ1 , . . . , θt . Then (M −
θ1 I) · · · (M − θt I) = 0, so that M t is a linear combination of I, M, . . . , M t−1 .
But if d(x, y) = t for two vertices x, y of Γ, then (M i )xy = 0 for 0 ≤ i ≤ t−1 and
(M t )xy > 0, contradiction. Hence t > d. This applies to M = A, to M = nI −L
and to M = Q, where A is the adjacency matrix, L is the Laplace matrix and
Q is the signless Laplace matrix of Γ.
Distance-regular graphs, discussed in Chapter 11, have equality here. For an
upper bound on the diameter, see §4.5.7.
1.3.5 Spanning trees

From the Laplace spectrum of a graph one can determine the number of spanning
trees (which will be nonzero only if the graph is connected).
1.3. THE SPECTRUM OF AN UNDIRECTED GRAPH 15
Proposition 1.3.4 Let Γ be an undirected (multi)graph with at least one vertex,

and Laplace matrix L with eigenvalues 0 = µ1 ≤ µ2 ≤ . . . ≤ µn . Let ℓxy be the
(x, y)-cofactor of L. Then the number N of spanning trees of Γ equals
1 1
N = ℓxy = det(L + J) = µ2 ...µn for any x, y ∈ V Γ.
n2 n
(The (i, j)-cofactor of a matrix M is by definition (−1)i+j det M (i, j), where
M (i, j) is the matrix obtained from M by deleting row i and column j. Note
that ℓxy does not depend on an ordering of the vertices of Γ.)
Proof. Let LS , for S ⊆ V Γ, denote the matrix obtained from L by deleting
the rows and columns indexed by S, so that ℓxx = det L{x} . The equality
N = ℓxx follows by induction on n, and for fixed n > 1 on the number of edges
incident with x. Indeed, if n = 1 then ℓxx = 1. Otherwise, if x has degree 0 then
ℓxx = 0 since L{x} has zero row sums. Finally, if xy is an edge, then deleting
this edge from Γ diminishes ℓxx by det L{x,y} , which by induction is the number
of spanning trees of Γ with edge xy contracted, which is the number of spanning
trees containing the edgeQxy. This shows N = ℓxx .
n
Now det(tI − L) = t i=2 (t − µi ) and (−1)n−1 µ2P ...µn is the coefficient of t,
d d {x}
that is, is P
dt det(tI − L)| t=0 . But dt det(tI − L) = x det(tI − L ) so that
µ2 ...µn = x ℓxx = nN .
Since the sum of the columns of L is zero, so that one column is minus
the sum of the other columns, we have ℓxx = ℓxy for any x, y. Finally, the
eigenvalues of L + n12 J are n1 and µ2 , . . . , µn , so det(L + n12 J) = n1 µ2 · · · µn .
For example, the multigraph of valency k on 2 vertices has Laplace matrix
k −k
L= so that µ1 = 0, µ2 = 2k, and N = 12 .2k = k.
−k k
If we consider the complete graph Kn , then µ2 = . . . = µn = n, and there-
fore Kn has N = nn−2 spanning trees. This formula is due to Cayley [78].
Proposition 1.3.4 is implicit in Kirchhoff [216] and known as the Matrix-Tree
Theorem.
There is a ‘1-line proof’ of the above result using the Cauchy-Binet formula.
Proposition 1.3.5 (Cauchy-Binet) Let A and B be m × n matrices. Then
X
det AB ⊤ = det AS det BS
S
n

where the sum is over the m-subsets S of the set of columns, and AS (BS )
m
is the square submatrix of order m of A (resp. B) with columns indexed by S.
2nd proof of Proposition 1.3.4 (sketch) Let Nx be the directed incidence

matrix of Γ, with row x deleted. Then lxx = det Nx Nx⊤ . Apply Cauchy-Binet
to get lxx as a sum of squares of determinants of size n − 1. These determinants
vanish unless the set S of columns is the set of edges of a spanning tree, in which
case the determinant is ±1.
1.3.6 Bipartite graphs

A graph Γ is called bipartite when its vertex set can be partitioned into two
disjoint parts X1 , X2 such that all edges of Γ meet both X1 and X2 . The

0 B
adjacency matrix of a bipartite graph has the form A = . It follows
B⊤ 0

u
that the spectrum of a bipartite graph is symmetric w.r.t. 0: if is an
v

u
eigenvector with eigenvalue θ, then is an eigenvector with eigenvalue −θ.
−v
(The converse also holds, see Proposition 3.4.1.)
One has rk A = 2 rk B. If ni = |Xi | (i = 1, 2) and n1 ≥ n2 , then rk A ≤ 2n2 ,
so that Γ has eigenvalue 0 with multiplicity at least n1 − n2 .
One cannot, in general, recognize bipartiteness from the Laplace or signless
Laplace spectrum. For example, K1,3 and K1 + K3 have the same signless
Laplace spectrum and only the former is bipartite. And Figure 13.4 gives an
example of a bipartite and non-bipartite graph with the same Laplace spectrum.
However, by Proposition 1.3.10 below, a graph is bipartite precisely when its
Laplace spectrum and signless Laplace spectrum coincide.
1.3.7 Connectedness
The spectrum of a disconnected graph is easily found from the spectra of its
connected components:
Proposition 1.3.6 Let Γ be a graph with connected components Γi (1 ≤ i ≤ s).

Then the spectrum of Γ is the union of the spectra of Γi (and multiplicities are
added). The same holds for the Laplace and the signless Laplace spectrum.
Proposition 1.3.7 The multiplicity of 0 as a Laplace eigenvalue of an undi-

rected graph Γ equals the number of connected components of Γ.
Proof. We have to show that a connected graph has Laplace eigenvalue 0

with multiplicity 1. As we saw earlier, L = N N ⊤ , where N is the incidence
matrix of an orientation of Γ. Now Lu = 0 is equivalent to N ⊤ u = 0 (since
0 = u⊤ Lu = ||N ⊤ u||2 ), that is, for every edge the vector u takes the same value
on both endpoints. Since Γ is connected that means that u is constant.
Proposition 1.3.8 Let the undirected graph Γ be regular of valency k. Then k

is the largest eigenvalue of Γ, and its multiplicity equals the number of connected
components of Γ.
Proof. We have L = kI − A.
One cannot see from the spectrum alone whether a (nonregular) graph is con-
nected: both K1,4 and K1 + C4 have spectrum 21 , 03 , (−2)1 (we write multiplic-
ities as exponents). And both Ê6 and K1 + C6 have spectrum 21 , 12 , 0, (−1)2 ,
(−2)1 .
Proposition 1.3.9 The multiplicity of 0 as a signless Laplace eigenvalue of an

undirected graph Γ equals the number of bipartite connected components of Γ.
Proof. Let M be the vertex-edge incidence matrix of Γ, so that Q = M M ⊤ .

If M M ⊤ u = 0 then M ⊤ u = 0, so that ux = −uy for all edges xy, and the
support of u is the union of a number of bipartite components of Γ.
1.4. SPECTRUM OF SOME GRAPHS 17
a
a a a a a a a
@ A
@a a a A a
a HH a
a a
@ A
a @a a a a HH a Aa a
K1,4 K1 + C 4 Ê6 K1 + C 6
Figure 1.1: Two pairs of cospectral graphs
Proposition 1.3.10 A graph Γ is bipartite if and only if the Laplace spectrum

and the signless Laplace spectrum of Γ are equal.
Proof. If Γ is bipartite, the Laplace matrix L and the signless Laplace ma-
trix Q are similar by a diagonal matrix D with diagonal entries ±1 (that is,
Q = DLD−1 ). Therefore Q and L have the same spectrum. Conversely, if
both spectra are the same, then by Propositions 1.3.7 and 1.3.9 the number of
connected components equals the number of bipartite components. Hence Γ is
bipartite.
1.4 Spectrum of some graphs

In this section we discuss some special graphs and their spectra. All graphs in
this section are finite, undirected and simple. Observe that the all-1 matrix J of
order n has rank 1, and that the all-1 vector 1 is an eigenvector with eigenvalue
n. So the spectrum is n1 , 0n−1 . (Here and throughout we write multiplicities
as exponents where that is convenient and no confusion seems likely.)
1.4.1 The complete graph

Let Γ be the complete graph Kn on n vertices. Its adjacency matrix is A = J −I,
and the spectrum is (n − 1)1 , (−1)n−1 . The Laplace matrix is nI − J, which
has spectrum 01 , nn−1 .
1.4.2 The complete bipartite graph

√
The spectrum of the complete bipartite graph Km,n is ± mn, 0m+n−2 . The
Laplace spectrum is 01 , mn−1 , nm−1 , (m + n)1 .
1.4.3 The cycle

Let Γ be the directed n-cycle Dn . Eigenvectors are (1, ζ, ζ 2 , . . . , ζ n−1 )⊤ where
ζ n = 1, and the corresponding eigenvalue is ζ. Thus, the spectrum consists
precisely of the complex n-th roots of unity e2πij/n (j = 0, . . . , n − 1).
Now consider the undirected n-cycle Cn . If B is the adjacency matrix of Dn ,
then A = B + B ⊤ is the adjacency matrix of Cn . We find the same eigenvectors
as before, with eigenvalues ζ +ζ −1 , so that the spectrum consists of the numbers
2 cos(2πj/n) (j = 0, . . . , n − 1).
This graph is regular of valency 2, so the Laplace spectrum consists of the
numbers 2 − 2 cos(2πj/n) (j = 0, . . . , n − 1).
1.4.4 The path

Let Γ be the undirected path Pn with n vertices. The ordinary spectrum consists
of the numbers 2 cos(πj/(n + 1)) (j = 1, . . . , n). The Laplace spectrum is
2 − 2 cos(πj/n) (j = 0, . . . , n − 1).
The ordinary spectrum follows by looking at C2n+2 . If u(ζ) = (1, ζ, ζ 2 , . . . ,
2n+1 ⊤
ζ ) is an eigenvector of C2n+2 , where ζ 2n+2 = 1, then u(ζ) and u(ζ −1 ) have
the same eigenvalue 2 cos(πj/(n + 1)), and hence so has u(ζ) − u(ζ −1 ). This
latter vector has two zero coordinates distance n + 1 apart and (for ζ 6= ±1)
induces an eigenvector on the two paths obtained by removing the two points
where it is zero.
Eigenvectors of L with eigenvalue 2 − ζ − ζ −1 are (1 + ζ 2n−1 , . . . , ζ j +
2n−1−j
ζ , . . . , ζ n−1 + ζ n ) where ζ 2n = 1. One can check this directly, or con-
sider Pn the result of folding C2n where the folding has no fixed vertices. An
eigenvector of C2n that is constant on the preimages of the folding yields an
eigenvector of Pn with the same eigenvalue.
1.4.5 Line graphs

The line graph L(Γ) of Γ is the graph with the edge set of Γ as vertex set, where
two vertices are adjacent if the corresponding edges of Γ have an endpoint in
common. If N is the incidence matrix of Γ, then N ⊤ N − 2I is the adjacency
matrix of L(Γ). Since N ⊤ N is positive semidefinite, the eigenvalues of a line
graph are not smaller than −2. We have an explicit formula for the eigenvalues
of L(Γ) in terms of the signless Laplace eigenvalues of Γ.
Proposition 1.4.1 Suppose Γ has m edges, and let ρ1 ≥ . . . ≥ ρr be the positive
signless Laplace eigenvalues of Γ, then the eigenvalues of L(Γ) are θi = ρi − 2
for i = 1, . . . , r, and θi = −2 if r < i ≤ m.
Proof. The signless Laplace matrix Q of Γ, and the adjacency matrix B of
L(Γ) satisfy Q = N N ⊤ and B + 2I = N ⊤N . Because N N ⊤ and N ⊤N have the
same nonzero eigenvalues (multiplicities included), the result follows.
Example Since the path Pn has line graph Pn−1 , and is bipartite, the Laplacian
and the signless Laplacian eigenvalues of Pn are 2 + 2 cos πi
n , i = 1, . . . , n.
Corollary 1.4.2 If Γ is a k-regular graph (k ≥ 2) with n vertices, e = kn/2
edges and eigenvalues θi (i = 1, . . . , n), then L(Γ) is (2k − 2)-regular with eigen-
values θi + k − 2 (i = 1, . . . , n) and e − n times −2.
The line graph of the complete graph Kn (n ≥ 2), is known as the triangular
graph T (n). It has spectrum 2(n − 2)1 , (n − 4)n−1 , (−2)n(n−3)/2 . The line graph
of the regular complete bipartite graph Km,m (m ≥ 2), is known as the lattice
2
graph L2 (m). It has spectrum 2(m − 1)1 , (m − 2)2m−2 , (−2)(m−1) . These
two families of graphs, and their complements, are examples of strongly regular
graphs, which will be the subject of Chapter 8. The complement of T (5) is the
famous Petersen graph. It has spectrum 31 15 (−2)4 .
1.4.6 Cartesian products

Consider the Cartesian product Γ × ∆ of two graphs Γ and ∆, where the vertex
set is the Cartesian product of the vertex sets of the factors, and two vertices
are adjacent where they agree in one coordinate and are adjacent in the other.
1.5. DECOMPOSITIONS 19
If u and v are eigenvectors for Γ and ∆ with ordinary or Laplace eigenvalues θ

and η, respectively, then the vector w defined by w(x,y) = ux vy is an eigenvector
of Γ × ∆ with ordinary or Laplace eigenvalue θ + η.
For example, L2 (m) = Km × Km .
For example, the hypercube 2n , also called Qn , is the Cartesian product of n
n
factors K2 . The spectrum of K2 is 1, −1, and hence the spectrum of 2 consists
of the numbers n − 2i with multiplicity ni (i = 0, 1, . . . , n).
1.4.7 Kronecker products and bipartite double

Given graphs Γ and ∆ with vertex sets V and W , respectively, their Kronecker
product (sometimes called conjunction) Γ⊗∆ is the graph with vertex set V ×W ,
where (v, w) ∼ (v ′ , w′ ) when v ∼ v ′ and w ∼ w′ . The adjacency matrix of
Γ ⊗ ∆ is the Kronecker product of the adjacency matrices of Γ and ∆.
If u and v are eigenvectors for Γ and ∆ with eigenvalues θ and η, respectively,
then the vector w = u ⊗ v (with w(x,y) = ux vy ) is an eigenvector of Γ ⊗ ∆ with
eigenvalue θη. Thus, the spectrum of Γ ⊗ ∆ consists of the products of the
eigenvalues of Γ and ∆.
Given a graph Γ, its bipartite double is the graph Γ ⊗ K2 (with for each
vertex x of Γ two vertices x′ and x′′ , and for each edge xy of Γ two edges x′ y ′′
and x′′ y ′ ). If Γ is bipartite, its double is just the union of two disjoint copies.
If Γ is connected and not bipartite, then its double is connected and bipartite.
If Γ has spectrum Φ, then Γ ⊗ K2 has spectrum Φ ∪ −Φ.
1.4.8 Cayley graphs

Let G be an abelian group and S ⊆ G. The Cayley graph on G with difference
set S is the (directed) graph Γ with vertex set G and edge set E = {(x, y) |
y − x ∈ S}. Now Γ is regular with in- and outvalency |S|. The graph Γ will be
undirected when S = −S.
It is easy to compute the spectrum of finite Cayley graphs. Let χ be a
∗
character
P of G, that is,P a map χ : G → C such that χ(x + y) = χ(x)χ(y).
Then y ∼ x χ(y) = ( s∈S χ(s))χ(x) so that the vector (χ(x))x∈GPis a right
eigenvector of the adjacency matrix A of Γ with eigenvalue χ(S) := s∈S χ(s).
The n = |G| distinct characters give independent eigenvectors, so one obtains
the entire spectrum in this way.
For example, the directed pentagon (with in- and outvalency 1) is a Cayley
graph for G = Z5 and S = {1}. The characters of G are the maps i 7→ ζ i
for some fixed 5-th root of unity ζ. Hence the directed pentagon has spectrum
{ζ | ζ 5 = 1}.
The undirected pentagon (with valency 2) is the Cayley graph for G = Z5
and S = {−1, 1}. The spectrum √ of the pentagon becomes {ζ + ζ −1 | ζ 5 = 1},
1
that is, consists of 2 and 2 (−1 ± 5) (both with multiplicity 2).
1.5 Decompositions
Here we present two non-trivial applications of linear algebra to graph decom-
positions.
1.5.1 Decomposing K10 into Petersen graphs

An amusing application ([30, 269]) is the following. Can the edges of the com-
plete graph K10 be colored with three colors such that each color induces a graph
isomorphic to the Petersen graph? K10 has 45 edges, 9 on each vertex, and the
Petersen graph has 15 edges, 3 on each vertex, so at first sight this might seem
possible. Let the adjacency matrices of the three color classes be P1 , P2 and P3 ,
so that P1 +P2 +P3 = J −I. If P1 and P2 are Petersen graphs, they both have a
5-dimensional eigenspace for eigenvalue 1, contained in the 9-space 1⊥ . There-
fore, there is a common 1-eigenvector u and P3 u = (J − I)u − P1 u − P2 u = −3u
so that u is an eigenvector for P3 with eigenvalue −3. But the Petersen graph
does not have eigenvalue −3, so the result of removing two edge-disjoint Pe-
tersen graphs from K10 is not a Petersen graph. (In fact, it follows that P3 is
connected and bipartite.)
1.5.2 Decomposing Kn into complete bipartite graphs

A famous result is the fact that for any edge-decomposition of Kn into complete
bipartite graphs one needs to use at least n − 1 summands. Since Kn has
eigenvalue −1 with multiplicity n − 1 this follows directly from the following:
Proposition 1.5.1 (H. S. Witsenhausen; Graham & Pollak [162]) Suppose a

graph Γ with adjacency matrix A has an edge decomposition into r complete
bipartite graphs. Then r ≥ n+ (A) and r ≥ n− (A), where n+ (A) and n− (A) are
the numbers of positive (negative) eigenvalues of A.
Proof. Let the i-th complete bipartite graph have a bipartition where ui and
vi are the characteristic vectors of both sides of the
P bipartition, so that its
adjacency matrix is Di = ui vi⊤ + vi u⊤i , and A = Di . Let w be a vector
orthogonal to all ui . Then w⊤ Aw = 0 and it follows that w cannot be chosen
in the span of eigenvectors of A with positive (negative) eigenvalue.
1.6 Automorphisms
An automorphism of a graph Γ is a permutation π of its point set X such that
x ∼ y if and only if π(x) ∼ π(y). Given π, we have a linear transformation Pπ
on V defined by (Pπ (u))x = uπ(x) for u ∈ V , x ∈ X. That π is an automorphism
is expressed by APπ = Pπ A. It follows that Pπ preserves the eigenspace Vθ for
each eigenvalue θ of A.
More generally, if G is a group of automorphisms of Γ then we find a linear
representation of degree m(θ) = dim Vθ of G.
We denote the group of all automorphisms of Γ by Aut Γ. One would expect
that when Aut Γ is large, then m(θ) tends to be large, so that Γ has only few
distinct eigenvalues. And indeed the arguments below will show that a transitive
group of automorphisms does not go very well together with simple eigenvalues.
Suppose dim Vθ = 1, say Vθ = hui. Since Pπ preserves Vθ we must have
Pπ u = ±u. So either u is constant on the orbits of π, or π has even order,
Pπ (u) = −u, and u is constant on the orbits of π 2 . For the Perron-Frobenius
eigenvector (cf. §2.2) we are always in the former case.
1.7. ALGEBRAIC CONNECTIVITY 21
Corollary 1.6.1 If all eigenvalues are simple, then Aut Γ is an elementary

abelian 2-group.
Proof. If π has order larger than two, then there are two distinct vertices x,
y in an orbit of π 2 , and all eigenvectors have identical x- and y-coordinates, a
contradiction.
Corollary 1.6.2 Let Aut Γ be transitive on X. (Then Γ is regular of degree k,

say.)
(i) If m(θ) = 1 for some eigenvalue θ 6= k, then v = |X| is even, and θ ≡ k

(mod 2). If Aut Γ is moreover edge-transitive then Γ is bipartite and
θ = −k.
(ii) If m(θ) = 1 for two distinct eigenvalues θ 6= k, then v ≡ 0 (mod 4).
(iii) If m(θ) = 1 for all eigenvalues θ, then Γ has at most two vertices.
Proof. (i) Suppose Vθ = hui. Then u induces a partition of X into two equal
parts: X = X+ ∪ X− , where ux = a for x ∈ X+ and ux = −a for x ∈ X− . Now
θ = k − 2|Γ(x) ∩ X− | for x ∈ X+ .
(ii) If m(k) = m(θ) = m(θ′ ) = 1, then we find 3 pairwise orthogonal (±1)-
vectors, and a partition of X into four equal parts.
(iii) There are not enough integers θ ≡ k (mod 2) between −k and k.
For more details, see Cvetković, Doob & Sachs [106], Chapter 5.
1.7 Algebraic connectivity

Let Γ be a graph with at least two vertices. The second smallest Laplace
eigenvalue µ2 (Γ) is called the algebraic connectivity of the graph Γ. This concept
was introduced by Fiedler [141]. Now, by Proposition 1.3.7, µ2 (Γ) ≥ 0, with
equality if and only if Γ is disconnected.
The algebraic connectivity is monotone: it does not decrease when edges are
added to the graph:
Proposition 1.7.1 Let Γ and ∆ be two edge-disjoint graphs on the same vertex
set, and Γ ∪ ∆ their union. We have µ2 (Γ ∪ ∆) ≥ µ2 (Γ) + µ2 (∆) ≥ µ2 (Γ).
Proof. Use that µ2 (Γ) = minu {u⊥ Lu | (u, u) = 1, (u, 1) = 0}.

The algebraic connectivity is a lower bound for the vertex connectivity:
Proposition 1.7.2 Let Γ be a graph with vertex set X. Suppose D ⊂ X is a

set of vertices such that the subgraph induced by Γ on X \ D is disconnected.
Then |D| ≥ µ2 (Γ).
Proof. By monotonicity we may assume that Γ contains all edges between

D and X \ D. Now a nonzero vector u that is 0 on D and constant on each
component of X \ D and satisfies (u, 1) = 0, is a Laplace eigenvector with
Laplace eigenvalue |D|.
1.8 Cospectral graphs

As noted above (in §1.3.7), there exist pairs of nonisomorphic graphs with the
same spectrum. Graphs with the same (adjacency) spectrum are called cospec-
tral (or isospectral). The two graphs of Figure 1.2 below are nonisomorphic and
cospectral. Both graphs are regular, which means that they are also cospectral
for the Laplace matrix, and any other linear combination of A, I, and J, includ-
ing the Seidel matrix (see §1.8.2), and the adjacency matric of the complement.
Q @
u u u u u u
A
@ @ @Q
A@u @u @uQ @u
A A Q A
AP uA u Q
QuA
@
uP
PP A
@ @A
u @u PP
Au
u @u @
Au
Figure 1.2: Two cospectral√regular graphs

√
(Spectrum: 4, 1, (−1)4 , ± 5, 12 (1 ± 17))
Let us give some more examples and families of examples. A more extensive
discussion is found in Chapter 13.
1.8.1 The 4-cube

The hypercube 2n is determined by its spectrum for n < 4, but not for n ≥ 4.
Indeed, there are precisely two graphs with spectrum 41 , 24 , 06 , (−2)4 , (−4)1
(Hoffman [194]). Consider the two binary codes of word length 4 and dimension
3 given by C1 = 1⊥ and C2 = (0111)⊥ . Construct a bipartite graph, where
one class of the bipartition consists of the pairs (i, x) ∈ {1, 2, 3, 4} × {0, 1} of
coordinate position and value, and the other class of the bipartition consists of
the code words, and code word u is adjacent to the pairs (i, ui ) for i ∈ {1, 2, 3, 4}.
For the code C1 this yields the 4-cube (tesseract), and for C2 we get its unique
cospectral mate.
1:1
• •
1110 1101 1011 1000

• • • • • • • •
• • • • • • • • • • • •
2:0 3:0 4:1 4:0 3:1 2:1
• • • • • • • •
0000 0011 0101 0110
• •
1:0
Figure 1.3: Tesseract and cospectral switched version

1.8. COSPECTRAL GRAPHS 23
1.8.2 Seidel switching

The Seidel adjacency matrix of a graph Γ with adjacency matrix A is the matrix
S defined by 
 0 if u = v
Suv = −1 if u ∼ v

1 if u 6∼ v
so that S = J − I − 2A. The Seidel spectrum of a graph is the spectrum of its
Seidel adjacency matrix. For a regular graph on n vertices with valency k and
other eigenvalues θ, the Seidel spectrum consists of n − 1 − 2k and the values
−1 − 2θ.
Let Γ have vertex set X, and let Y ⊂ X. Let D be the diagonal matrix
indexed by X with Dxx = −1 for x ∈ Y , and Dxx = 1 otherwise. Then
DSD has the same spectrum as S. It is the Seidel adjacency matrix of the
graph obtained from Γ by leaving adjacency and nonadjacency inside Y and
X \ Y as it was, and interchanging adjacency and nonadjacency between Y and
X \Y . This new graph, Seidel-cospectral with Γ, is said to be obtained by Seidel
switching with respect to the set of vertices Y .
Being related by Seidel switching is an equivalence relation, and the equiv-
alence classes are called switching classes. Here are the three switching classes
of graphs with 4 vertices.
s s s s s s s s s s s s s s s s s s s s s s
s s∼ s s ∼ s s / s s ∼ s s∼ s s ∼ s s ∼ s s / s s∼ s s ∼ @
s s
The Seidel matrix of the complementary graph Γ is −S, so that a graph and
its complement have opposite Seidel eigenvalues.
If two regular graphs of the same valency are Seidel cospectral, then they
are also cospectral.
Figure 1.2 shows an example of two cospectral graphs related by Seidel
switching (with respect to the four corners). These graphs are nonisomorphic:
they have different local structure.
The Seidel adjacency matrix plays a rôle in the description of regular two-
graphs (see §§9.1–9.3) and equiangular lines (see §9.6).
1.8.3 Godsil-McKay switching

Let Γ be a graph with vertex set X, and let {C1 , . . . , Ct , D} be a partition of
X such that {C1 , . . . , Ct } is an equitable partition of X \ D (that is, any two
vertices in Ci have the same number of neighbors in Cj for all i, j), and for every
x ∈ D and every i ∈ {1, . . . , t} the vertex x has either 0, 21 |Ci | or |Ci | neighbors
in Ci . Construct a new graph Γ′ by interchanging adjacency and nonadjacency
between x ∈ D and the vertices in Ci whenever x has 12 |Ci | neighbors in Ci .
Then Γ and Γ′ are cospectral ([156]).
2
Indeed, let Qm be the matrix m J − I of order m, so that Q2m = I. Let
ni = |Ci |. Then the adjacency matrix A′ of Γ′ is found to be QAQ where Q is
the block diagonal matrix with blocks Qni (1 ≤ i ≤ t) and I (of order |D|).
The same argument also applies to the complementary graphs, so that also
the complements of Γ and Γ′ are cospectral. Thus, for example, the second pair
of graphs in Figure 1.1 is related by GM-switching, and hence has cospectral

complements. The first pair does not have cospectral complements and hence
does not arise by GM-switching.
The 4-cube and its cospectral mate (Figure 1.3) can be obtained from each
other by GM-switching with respect to the neighborhood of a vertex. Figure
1.2 is also an example of GM-switching. Indeed, when two regular graphs of
the same degree are related by Seidel switching, the switch is also a case of
GM-switching.
1.8.4 Reconstruction
The famous Kelly-Ulam conjecture (1941) asks whether a graph Γ can be re-
constructed when the (isomorphism types of) the n vertex-deleted graphs Γ \ x
are given. The conjecture is still open (see Bondy [29] for a discussion), but
Tutte [297] showed that one can reconstruct the characteristic polynomial of Γ,
so that any counterexample to the reconstruction conjecture must be a pair of
cospectral graphs.
1.9 Very small graphs

Let us give various spectra for the graphs on at most 4 vertices. The columns
with heading A, L, Q, S give the spectrum for the adjacency matrix, the Laplace
matrix L = D − A (where D is the diagonal matrix of degrees), the signless
Laplace matrix Q = D + A and the Seidel matrix S = J − I − 2A.
label picture A L Q S
0.1
1.1 s 0 0 0 0
2.1 s s 1, −1 0, 2 2, 0 −1, 1
2.2 s s 0, 0 0, 0 0, 0 −1, 1
s
3.1 sAs 2, −1, −1 0, 3, 3 4, 1, 1 −2, 1, 1
s √ √
3.2 A
s s 2, 0, − 2 0, 1, 3 3, 1, 0 −1, −1, 2
s
3.3 s s 1, 0, −1 0, 0, 2 2, 0, 0 −2, 1, 1
s
3.4 s s 0, 0, 0 0, 0, 0 0, 0, 0 −1, −1, 2
s s
4.1 @
s s 3, −1, −1, −1 0, 4, 4, 4 6, 2, 2, 2 −3, 1, 1, 1
s s √ √
4.2 s s ρ, 0, −1, 1−ρ 0, 2, 4, 4 2+2τ, 2, 2, 4−2τ − 5, −1, 1, 5
s s
4.3 s s 2, 0, 0, −2 0, 2, 2, 4 4, 2, 2, 0 −1, −1, −1, 3
s s √ √
4.4 s s θ1 , θ2 , −1, θ3 0, 1, 3, 4 2+ρ, 2, 1, 3−ρ − 5, −1, 1, 5
s s √ √
4.5 s s 3, 0, 0, − 3 0, 1, 1, 4 4, 1, 1, 0 −1, −1, −1, 3
s s √ √
4.6 s s τ, τ −1, 1−τ, −τ 0, 4−α, 2, α α, 2, 4−α, 0 − 5, −1, 1, 5
s s
4.7 s s 2, 0, −1, −1 0, 0, 3, 3 4, 1, 1, 0 −3, 1, 1, 1
continued...
1.10. EXERCISES 25
label picture A L Q S
s s √ √ √ √
4.8 s s 2, 0, 0, − 2 0, 0, 1, 3 3, 1, 0, 0 − 5, −1, 1, 5
s s
4.9 s s 1, 1, −1, −1 0, 0, 2, 2 2, 2, 0, 0 −3, 1, 1, 1
s s √ √
4.10 s s 1, 0, 0, −1 0, 0, 0, 2 2, 0, 0, 0 − 5, −1, 1, 5
s s
4.11 s s 0, 0, 0, 0 0, 0, 0, 0 0, 0, 0, 0 −1, −1, −1, 3
Table 1.1: Spectra of very small graphs

√ √ √
Here α = 2 + 2 and τ = (1 + 5)/2 and ρ = (1 + 17)/2 and θ1 ≈ 2.17009,
θ2 ≈ 0.31111, θ3 ≈ −1.48119 are the three roots of θ3 − θ2 − 3θ + 1 = 0.
1.10 Exercises
Exercise 1 Show thatpno graph has eigenvalue −1/2. Show that no undirected
√
graph has eigenvalue 2 + 5. (Hint: consider the algebraic conjugates of this
number.)
Exercise 2 Let Γ be an undirected graph with eigenvalues θ1 , . . . , θn . Show
that for any two vertices a and b of Γ there are constants
P h c1 , . . . , cn such that
the number of walks of length h from a to b equals ci θi for all h.
Exercise 3 Let Γ be a directed graph with constant outdegree k > 0 and
without directed 2-cycles. Show that Γ has a non-real eigenvalue.
Exercise 4 (i) Let Γ be a directed graph on n vertices, such that there is a
h with the property that for any two vertices a and b (distinct or not) there
is a unique directed path of length h from a to b. Prove that Γ has constant
in-degree and out-degree k, where n = k h , and has spectrum k 1 0n−1 .
(ii) The de Bruijn graph of order m is the directed graph with as vertices
the 2m binary sequences of length m, where there is an arrow from a1 . . . am to
b1 . . . bm when the tail a2 . . . am of the first equals the head b1 . . . bm−1 of the
second. (For m = 0 a single vertex with two loops.) Determine the spectrum of
the de Bruijn graph.
(iii) A de Bruijn cycle of order m ≥ 1 ([64, 65, 145]) is a circular arrangement
of 2m zeros and ones such that each binary sequence of length m occurs once in
this cycle. (In other words, it is a Hamiltonian cycle in the de Bruijn graph of
order m, an Eulerian cycle in the de Bruijn graph of order m − 1.) Show that
m−1
there are precisely 22 −m
de Bruijn cycles of order m.
Exercise 5 ([37, 265]) Let Γ be a tournament, that is, a directed graph in which
there is precisely one edge between any two distinct vertices, in other words, of
which the adjacency matrix A satisfies A⊤ + A = J − I.
(i) Show that all eigenvalues have real part not less than −1/2.
(ii) The tournament Γ is called transitive if (x, z) is an edge whenever both

(x, y) and (y, z) are edges. Show that all eigenvalues of a transitive tour-
nament are zero.
(iii) The tournament Γ is called regular when each vertex has the same number
of out-arrows. Clearly, when there are n vertices, this number of out-
arrows is (n − 1)/2. Show that all eigenvalues θ have real part at most
(n − 1)/2, and that Re(θ) = (n − 1)/2 occurs if and only if Γ is regular
(and then θ = (n − 1)/2).
(iv) Show that A either has full rank n or has rank n − 1, and that A has full
rank when Γ is regular and n > 1.
(Hint: for a vector u, consider the expression ū⊤ (A⊤ + A)u.)

Exercise 6 Let Γ be bipartite and consider its line graph L(Γ).
(i) Show that Γ admits a directed incidence matrix N such that N ⊤N − 2I

is the adjacency matrix of L(Γ).
(ii) Give a relation between the Laplace eigenvalues of Γ and the ordinary
eigenvalues of L(Γ).
(iii) Verify this relation in case Γ is the path Pn .
Exercise 7 ([94]) Verify (see §1.2.1) that both graphs pictured here have char-
acteristic polynomial t4 (t4 − 7t2 + 9), so that these two trees are cospectral.
s
s s
s
s@s s s s s s s

Hs
s @s JJs
Note how the coefficients of the characteristic polynomial of a tree count

partial matchings (sets of pairwise disjoint edges) in the tree.
Exercise 8 ([14]) Verify that both graphs pictured here have characteristic
polynomial (t − 1)(t + 1)2 (t3 − t2 − 5t + 1) by computing eigenvectors and
eigenvalues. Use the observation (§1.6) that the image of an eigenvector under
an automorphism is again an eigenvector. In particular, when two vertices x, y
are interchanged by an involution (automorphism of order 2), then a basis of
the eigenspace exists consisting of vectors where the x- and y-coordinates are
either equal or opposite.
s s s
s s H H
s
s
Hs s s sHs
Exercise 9 Let the cone over a graph Γ be the graph obtained by adding
a new vertex and joining that to all vertices of Γ. If Γ is regular of valency
k on n vertices, then show that the cone over Γ has characteristic polynomial
p(x) = (x2 − kx − n)pΓ (x)/(x − k).
Exercise 10 Show that the Seidel adjacency matrix S of a graph on n vertices
has rank n − 1 or n. (Hint: det S ≡ n − 1 (mod 2).)
Exercise 11 Prove that the complete graph K55 is not the union of three
copies of the triangular graph T (11).
Chapter 2
Linear algebra
In this chapter we present some less elementary, but relevant results from linear
algebra.
2.1 Simultaneous diagonalization

Let V be a complex vector space with finite dimension, andP fix a basis. Then
we can define an inner product on V by putting (x, y) = xi yi = x⊤ y for
x, y ∈ V , where the bar denotes complex conjugation. If A is Hermitean, i.e.,
if (Ax, y) = (x, Ay) for all x, y ∈ V , then all eigenvalues of A are real, and V
admits an orthonormal basis of eigenvectors of A.
Proposition 2.1.1 Suppose A is a collection of commuting Hermitean linear

transformations on V (i.e., AB = BA for A, B ∈ A), then V has a basis
consisting of common eigenvectors of all A ∈ A.
Proof. Induction on dim V . If each A ∈ A is a multiple of the identity I,

then all is clear. Otherwise, let A ∈ A not be a multiple of I. If Au = θu and
B ∈ A, then A(Bu) = BAu = θBu so that B acts as a linear transformation
on the eigenspace Vθ for the eigenvalue θ of A. By the induction hypothesis we
can choose a basis consisting of common eigenvectors for each B ∈ A in each
eigenspace. The union of these bases is the basis of V we were looking for.
Given a square matrix A, we can regard A as a linear transformation on a vector
space (with fixed basis). Hence the above concepts apply. The matrix A will
be Hermitean precisely when A = Ā⊤ ; in particular, a real symmetric matrix is
Hermitean.
2.2 Perron-Frobenius Theory

Let T be a real n × n matrix with nonnegative entries. T is called primitive if
for some k we have T k > 0; T is called irreducible if for all i, j there is a k such
that (T k )ij > 0.
Here, for a matrix (or vector) A, A > 0 (≥ 0) means that all its entries are
positive (nonnegative).
27
28 CHAPTER 2. LINEAR ALGEBRA
The matrix T = (tij ) is irreducible if and only if the directed graph ΓT with
vertices {1, . . . , n} and edges (i, j) whenever tij > 0 is strongly connected.
(A directed graph (X, E) is strongly connected if for any two vertices x, y
there is a directed path from x to y, i.e., there are vertices x0 = x, x1 , . . . , xm = y
such that (xi−1 , xi ) ∈ E for 1 ≤ i ≤ m.)
Note that if T is irreducible, then I + T is primitive.
The period d of an irreducible matrix T is the greatest common divisor of
the integers k for which (T k )ii > 0. It is independent of the i chosen.
Theorem 2.2.1 Let T ≥ 0 be irreducible. Then there is a (unique) positive

real number θ0 with the following properties:
(i) There is a real vector x0 > 0 with T x0 = θ0 x0 .
(ii) θ0 has geometric and algebraic multiplicity one.
(iii) For each eigenvalue θ of T we have |θ| ≤ θ0 . If T is primitive, then
|θ| = θ0 implies θ = θ0 . In general, if T has period d, then T has precisely
d eigenvalues θ with |θ| = θ0 , namely θ = θ0 e2πij/d for j = 0, 1, ..., d − 1.
In fact the entire spectrum of T is invariant under rotation of the complex
plane over an angle 2π/d about the origin.
(iv) Any nonnegative left or right eigenvector of T has eigenvalue θ0 . More
generally, if x ≥ 0, x 6= 0 and T x ≤ θx, then x > 0 and θ ≥ θ0 ; moreover,
θ = θ0 if and only if T x = θx.
(v) If 0 ≤ S ≤ T or if S is a principal minor of T , and S has eigenvalue σ,
then |σ| ≤ θ0 ; if |σ| = θ0 , then S = T .
(vi) Given a complex matrix S, let |S| denote the matrix with elements |S|ij =
|Sij |. If |S| ≤ T and S has eigenvalue σ, then |σ| ≤ θ0 . If equality
holds, then |S| = T , and there are a diagonal matrix E with diagonal
entries of absolute value 1 and a constant c of absolute value 1, such that
S = cET E −1 .
Proof. (i) Let P = (I + T )n−1 . Then P > 0 and P T = T P . Let B = {x |

x ≥ 0 and x 6= 0}. Define for x ∈ B:
(T x)i
θ(x) = max {θ | θ ∈ R, θx ≤ T x} = min { | 1 ≤ i ≤ n, xi 6= 0}.
xi
Now θ(αx) = θ(x) for α ∈ R, α > 0, and (x ≤ y, x 6= y implies P x θ(x) unless x is an eigenvector of T . Put
C = {x | x ≥ 0 and kxk = 1}. Then since C is compact and θ(.) is continuous
on P [C] (but not in general on C !), there is an x0 ∈ P [C] such that
θ0 := sup θ(x) = sup θ(x) = sup θ(x) = θ(x0 ).

x∈B x∈C x∈P [C]
Now x0 > 0 and x0 is an eigenvector of T , so T x0 = θ0 x0 , and θ0 > 0.

(ii) For a vector x = (x1 , ..., xn )⊤ , write x+ = (|x1 |, ..., |xn |)⊤ . If T x = θx,
then by the triangle inequality we have T x+ ≥ |θ|x+ . For nonzero x this means
|θ| ≤ θ(x+ ) ≤ θ0 . If, for some vector z ∈ B, we have T z ≥ θ0 z, then z is
eigenvector of T (otherwise θ(P z) > θ0 ), and since 0 0. Now if x is any eigenvector of T with eigenvalue θ0 , then if x is

real consider y = x0 + εx, where ε is chosen such that y ≥ 0 but not y > 0; now
y 6∈ B, i.e., y = 0, so x is a multiple of x0 . In the general case, both the real
and imaginary parts of x are multiples of x0 . This shows that the eigenspace
of θ0 has dimension 1, i.e., that the geometric multiplicity of θ0 is 1. We shall
look at the algebraic multiplicity later.
(iii) We have seen |θ| ≤ θ0 . If |θ| = θ0 and P T x = θx, then
P T x+ = θ0 x+ and
we had equality in the triangle inequality | j tij xj | ≤ j tij |xj |; this means
that all numbers tij xj (1 ≤ j ≤ n) have the same angular part (argument). If
T is primitive, then we can apply this reasoning with T k instead of T , where
T k > 0, and conclude that all xj have the same angular part. Consequently, in
this case x is a multiple of a real vector and may be taken real, nonnegative. Now
T x = θx shows that θ is real, and |θ| = θ0 that θ = θ0 . In the general case, T d
is a direct sum of primitive matrices T (0) , ..., T (d−1) , and if x = (x(0) , ..., x(d−1) )
is the corresponding decomposition of an eigenvector of T (with eigenvalue θ),
then (x(0) , ζx(1) , ..., ζ d−1 x(d−1) ) also is an eigenvector of T , with eigenvalue ζθ,
for any d-th root of unity ζ. (Here we assume that the T (i) are ordered in such
a way that in ΓT the arrows point from the subset corresponding to T (i) to the
subset corresponding to T (i+1) .) Since T d has a unique eigenvalue of maximum
(i+1)
modulus (let T(i) be the (nonsquare) submatrix of T describing the arrows
in ΓT between the subset corresponding to T (i) to the subset corresponding to
Qd−1 (i+j+1)
T (i+1) ; then T (i) = j=0 T(i+j) and if T (i) z = γz, z > 0 then T (i−1) z ′ = γz ′
(i)
where z ′ = T(i−1) z 6= 0, so that all T (i) have the same eigenvalue of maximum
modulus), it follows that T has precisely d such eigenvalues.
(iv) Doing the above for left eigenvectors instead of right ones, we find y0 > 0
with y0⊤ T = η0 y0⊤ . If T x = θx and y ⊤ T = ηy ⊤ , then ηy ⊤ x = y ⊤ T x = θy ⊤ x.
It follows that either θ = η or y ⊤ x = 0. Taking y ∈ B, x = x0 or x ∈ B,
y = y0 we see that θ = η (= θ0 = η0 ). Similarly, if T x ≤ θx, x ∈ B then
θ0 y0⊤ x = y0⊤ T x ≤ θy0⊤ x so that θ0 ≤ θ; also 0 0. If
θ = θ0 , then y0⊤ (T x − θx) = 0 so T x = θx.
(v) If s 6= 0, Ss = σs, then T s+ ≥ Ss+ ≥ |σ|s+ , so |σ| ≤ θ0 . But if |σ| = θ0
then s+ is eigenvector of T and s+ > 0 and (T − S)s+ = 0, so S = T .
(vi) If s 6= 0, Ss = σs, then T s+ ≥ |S|s+ ≥ |σ|s+ , so |σ| ≤ θ0 , and if |σ| = θ0
then s+ is eigenvector
P of TPand s+ > 0 and |S| = T . Equality in |S|s+ = |σ|s+
means that | Sij sj | = |Sij |.|sj |, so that given i all Sij sj have the same
−1
angular part. Let Eii = si /|si | and c = σ/|σ|. Then Sij = cEii Ejj |Sij |.
(vii) Finally, in order to prove that θ0 is a simple root of χT , the characteristic
d
polynomial of T , we have to show that dθ χT (θ) is nonzero for θ = θ0 . But
d
P
χT (θ) = det(θI − T ) and dθ χT (θ) = i det(θI − Tii ), and by (v) we have
det(θI − Tii ) > 0 for θ = θ0 .
Remark In case T ≥ 0 but T not necessarily irreducible, we can say the
following.
(i) The spectral radius θ0 of T is an eigenvalue, and there are nonnegative

left and right eigenvectors corresponding to it.
(ii) If |S| ≤ T and S has eigenvalue σ, then |σ| ≤ θ0 .
( Proof. (i) Use continuity arguments; (ii) the old proof still applies. )
For more details, see the exposition of the Perron-Frobenius theory in Gant-
macher [149], Chapter XIII; cf. also Varga [299], Marcus & Minc [238], Berman
& Plemmons [22], Seneta [279, Ch. 1], or Horn & Johnson [200, Ch. 8].
2.3 Equitable partitions

Suppose A is a symmetric real matrix whose rows and columns are indexed by
X = {1, . . . , n}. Let {X1 , . . . , Xm } be a partition of X. The characteristic
matrix S is the n × m matrix whose j th column is the characteristic vector of
Xj (j = 1, . . . , m). Define ni = |Xi | and K = diag(n1 , . . . , nm ). Let A be
partitioned according to {X1 , . . . , Xm }, that is
 
A1,1 . . . A1,m
 
A =  ... ..
. ,
Am,1 ... Am,m
wherein Ai,j denotes the submatrix (block) of A formed by rows in Xi and the
columns in Xj . Let bi,j denote the average row sum of Ai,j . Then the matrix
B = (bi,j ) is called the quotient matrix. We easily have
KB = S ⊤ AS, S ⊤ S = K.
If the row sum of each block Ai,j is constant then the partition is called equitable
(or regular) and we have Ai,j 1 = bi,j 1 for i, j = 0, . . . , d, so
AS = SB.
The following result is well-known and useful.

Lemma 2.3.1 If, for an equitable partition, v is an eigenvector of B for an
eigenvalue λ, then Sv is an eigenvector of A for the same eigenvalue λ.
Proof. Bv = θv implies ASv = SBv = θSv.
In the situation of this lemma, the spectrum of A consists of the spectrum
of the quotient matrix B (with eigenvectors in the column space of S, i.e.,
constant on the parts of the partition) together with the eigenvalues belonging
to eigenvectors orthogonal to the columns of S (i.e., summing to zero on each
part of the partition). These latter eigenvalues remain unchanged if the blocks
Ai,j are replaced by Ai,j + ci,j J for certain constants ci,j .
2.3.1 Equitable and almost equitable partitions of graphs

If in the above the matrix A is the adjacency matrix (or the Laplace matrix)
of a graph, then an equitable partition of the matrix A is a partition of the
vertex set into parts Xi such that each vertex in Xi has the same number bi,j
of neighbours in part Xj , for any j (or any j 6= i). Such partitions are called
(almost) equitable partitions of the graph.
For example, the adjacency matrix of the complete bipartite graph
Kp,q has an
0p
equitable partition with m = 2. The quotient matrix B equals and has
q0
√
eigenvalues ± pq, which are the nonzero eigenvalues of Kp,q .
2.4. THE RAYLEIGH QUOTIENT 31
More generally, consider the join Γ of two vertex-disjoint graphs Γ1 and Γ2 , the
graph obtained by inserting all possible edges between Γ1 and Γ2 . If Γ1 and Γ2
have n1 resp. n2 vertices and are both regular, say of valency k1 resp. k2 , and
have spectra Φ1 resp. Φ2 , then Γ has spectrum Φ = (Φ1 \ {k1 }) ∪ (Φ2 \ {k2 }) ∪
{k ′ , k ′′ } where k ′ , k ′′ are the two eigenvalues of

k1 n2
.
n1 k2
Indeed, we have an equitable partition of the adjacency matrix of Γ with the
above quotient matrix. The eigenvalues that do not belong to the quotient
coincide with those of the disjoint union of Γ1 and Γ2 .
2.4 The Rayleigh quotient

Let A be a real symmetric matrix and let u be a nonzero vector. The Rayleigh
quotient of u w.r.t. A is defined as
u⊤ Au
.
u⊤ u
Let u1 , . . . , un be an orthonormal
P set of eigenvectors P of A, say with Au
Pi =2 θi ui ,
where θ1 ≥ . . . ≥ θn . If u = αi ui then u⊤ u = αi2 and u⊤ Au = αi θi . It
follows that
u⊤ Au
≥ θi if u ∈ hu1 , . . . , ui i
u⊤ u
and
u⊤ Au ⊥
≤ θi if u ∈ hu1 , . . . , ui−1 i .
u⊤ u
In both cases, equality implies that u is a θi -eigenvector of A. Conversely, one
has
Theorem 2.4.1 (Courant-Fischer) Let W be an i-subspace of V . Then
u⊤ Au
θi ≥ min
u∈W, u6=0 u⊤ u
and
u⊤ Au
θi+1 ≤ max .
u∈W ⊥ , u6=0 u⊤ u
2.5 Interlacing
Consider two sequences of real numbers: θ1 ≥ . . . ≥ θn , and η1 ≥ . . . ≥ ηm with
m < n. The second sequence is said to interlace the first one whenever
θi ≥ ηi ≥ θn−m+i , for i = 1, . . . , m.
The interlacing is tight if there exist an integer k ∈ [0, m] such that
θi = ηi for 1 ≤ i ≤ k and θn−m+i = ηi for k + 1 ≤ i ≤ m.
If m = n − 1, the interlacing inequalities become θ1 ≥ η1 ≥ θ2 ≥ η2 ≥ . . . ≥
ηm ≥ θn , which clarifies the name. Godsil [154] reserves the name ‘interlacing’
for this particular case and calls it generalized interlacing otherwise.
Theorem 2.5.1 Let S be a real n×m matrix such that S ⊤ S = I. Let A be a real
symmetric matrix of order n with eigenvalues θ1 ≥ . . . ≥ θn . Define B = S ⊤ AS
and let B have eigenvalues η1 ≥ . . . ≥ ηm and respective eigenvectors v1 . . . vm .
(i) The eigenvalues of B interlace those of A.
(ii) If ηi = θi or ηi = θn−m+i for some i ∈ [1, m], then B has a ηi -eigenvector

v such that Sv is a ηi -eigenvector of A.
(iii) If for some integer l, ηi = θi , for i = 1, . . . , l (or ηi = θn−m+i for i =

l, . . . , m), then Svi is a ηi -eigenvector of A for i = 1, . . . , l (respectively
i = l, . . . , m).
(iv) If the interlacing is tight, then SB = AS.
Proof. Let u1 , . . . , un be an orthonormal set of eigenvectors of the matrix A,

where Aui = θi . For each i ∈ [1, m], take a nonzero vector si in

⊥
hv1 , . . . , vi i ∩ S ⊤ u1 , . . . , S ⊤ ui−1 . (2.1)
⊥
Then Ssi ∈ hu1 , . . . , ui−1 i , hence by Rayleigh’s principle,
⊤
(Ssi ) A(Ssi ) si ⊤ Bsi
θi ≥ ⊤
= ≥ ηi ,
(Ssi ) (Ssi ) si ⊤ si
and similarly (or by applying the above inequality to −A and −B) we get
θn−m+i ≤ ηi , proving (i). If θi = ηi , then si and Ssi are θi -eigenvectors of
B and A, respectively, proving (ii). We prove (iii) by induction on l. Assume
Svi = ui for i = 1, . . . , l − 1. Then we may take sl = vl in (2.1), but in proving
(ii) we saw that Ssl is a θl -eigenvector of A. (The statement between parentheses
follows by considering −A and −B.) Thus we have (iii). Let the interlacing be
tight. Then by (iii), Sv1 , . . . , Svm is an orthonormal set of eigenvectors of A for
the eigenvalues η1 , . . . , ηm . So we have SBvi = ηi Svi = ASvi , for i = 1, . . . , m.
Since the vectors vi form a basis, it follows that SB = AS.
If we take S = [I 0]⊤ , then B is just a principal submatrix of A and we have
the following corollary.
Corollary 2.5.2 If B is a principal submatrix of a symmetric matrix A, then

the eigenvalues of B interlace the eigenvalues of A.
The theorem requires the columns of S to be orthonormal. If one has a situ-

ation with orthogonal but not necessarily orthonormal vectors, some scaling is
required.
Corollary 2.5.3 Let A be a real symmetric matrix of order n. Let x1 , . . . , xm

be nonzero orthogonal real vectors of order n. Define a matrix C = (cij ) by
cij = ||x1i ||2 x⊤
i Axj .
(i) The eigenvalues of C interlace the eigenvalues of A.

P
(ii) If the interlacing is tight, then Axj = cij xi for all j.
2.6. SCHUR’S INEQUALITY 33
P ⊤
(iii) Let x = xj . The number r := xx⊤Ax
x
lies between the smallest and largest
eigenvalue of C. If x is an eigenvector of A with eigenvalue θ, then also
C has an eigenvalue θ (for eigenvector 1).
Proof. Let K be the diagonal matrix with Kii = ||xi ||. Let R be the n × m
matrix with columns xj , and put S = RK −1 . Then S ⊤ S = I, and the theorem
applies with B = S ⊤ AS = KCK −1 . If interlacing is tight we have AR = RC.
P ⊤ ⊤
With x = xj = R1 and y = K1, we have xx⊤Axx
= yy⊤Byy
.
In particular, this applies when the xi are the characteristic vectors of a partition
(or just a collection of pairwise disjoint subsets).
Corollary 2.5.4 Let C be the quotient matrix of a symmetric matrix A whose

rows and columns are partitioned according to a partitioning {X1 , . . . , Xm }.
(i) The eigenvalues of C interlace the eigenvalues of A.
(ii) If the interlacing is tight, then the partition is equitable.
Theorem 2.5.1(i) is a classical result; see Courant & Hilbert [98], Vol. 1, Ch. I.
For the special case of a principal submatrix (Corollary 2.5.2), the result even
goes back to Cauchy and is therefore often referred to as Cauchy interlacing.
Interlacing for the quotient matrix (Corollary 2.5.4) is especially applicable to
combinatorial structures (as we shall see). Payne (see, for instance, [256]) has
applied the extremal inequalities θ1 ≥ ηi ≥ θn to finite geometries several
times. He attributes the method to Higman and Sims and therefore calls it the
Higman-Sims technique.
Remark This theorem generalizes directly to complex Hermitean matrices in-
stead of real symmetric matrices (with conjugate transpose instead of transpose)
with virtually the same proof.
For more detailed eigenvalue inequalities, see Haemers [175], [177].
2.6 Schur’s inequality

Theorem 2.6.1 (Schur [266]) Let A be a real symmetric matrix with eigen-
values
Pt θ1 ≥Pθ2 ≥ . . . ≥ θn and diagonal elements d1 ≥ d2 ≥ . . . ≥ dn . Then
t
i=1 i ≤
d i=1 θi for 1 ≤ t ≤ n.
Proof. Let B be the principal submatrix of A obtained by deleting the rows

and columns containing
Pt dt+1 , . . .P has eigenvalues ηi (1 ≤ i ≤ t) then
, dn . If B P
t t
by interlacing i=1 di = tr B = i=1 ηi ≤ i=1 θi .
Remark Again ‘real symmetric’ can be replaced by ‘Hermitean’.
2.7 Schur complements

In this section, the square matrix

A11 A12
A=
A21 A22
is a square partitioned matrix (over any field), where A11 is nonsingular. The
Schur complement A/A11 of A11 in A is the matrix A22 − A21 A−1 11 A12 . The
following result is a straightforward but important consequence from the defi-
nition.
Theorem 2.7.1 (see [311]) The Schur complement A/A11 satisfies

I O A11 A12 I −A−1
11 A12 A11 O
(i) = ,
−A21 A−1
11 I A21 A22 O I O A/A11
(ii) det(A/A11 ) = det A/ det A11 ,
(iii) rk A = rk A11 + rk(A/A11 ).
Corollary 2.7.2 If rk A = rk A11 , then A22 = A21 A−1

11 A12 .
2.8 The Courant-Weyl inequalities

Denote the eigenvalues of a Hermitean matrix A, arranged in nonincreasing
order, by λi (A).
Theorem 2.8.1 Let A and B be Hermitean matrices of order n, and let 1 ≤

i, j ≤ n.
(i) If i + j − 1 ≤ n then λi+j−1 (A + B) ≤ λi (A) + λj (B).
(ii) If i + j − n ≥ 1 then λi (A) + λj (B) ≤ λi+j−n (A + B).
(iii) If B is positive semidefinite, then λi (A + B) ≥ λi (A).
Proof. (i) Let u1 , . . . , un and v1 , . . . , vn be orthonormal sets of eigenvectors of

A resp. B, with Aui = λi (A)ui and Bvj = λj (B)vj . Let U = huh | 1 ≤ h ≤ i−1i
and V = hvh | 1 ≤ h ≤ j − 1i, and W = U + V . For w ∈ W ⊥ we have
w⊤ (A + B)w ≤ (λi (A) + λj (B))w⊤ w. It follows that the space spanned by
eigenvectors of A + B with eigenvalue larger than λi (A) + λj (B) has dimension
at most i + j − 2.
(ii) Apply (i) to −A and −B. (iii) Apply the case j = n of (ii).
Ky Fan [138] shows that λ(A) + λ(B) dominates λ(A + B):
Theorem 2.8.2 Let A Pand B be Hermitean

Pt matrices of
Porder n. Then for all
t t
t, 0 ≤ t ≤ n, we have i=1 λi (A + B) ≤ i=1 λi (A) + i=1 λi (B).
Pt ∗
Proof. i=1 λi (A) = max tr(U AU ), where the maximum is over all n × t
∗
matrices U with U U = I.
2.9. GRAM MATRICES 35
2.9 Gram matrices

Real symmetric n × n-matrices G are in bijective correspondence with quadratic
forms q on Rn via the relation
q(x) = x⊤ Gx (x ∈ Rn ).
Two quadratic forms q and q ′ on Rn are congruent, i.e., there is a nonsingular

n × n-matrix S such that q(x) = q ′ (Sx) for all x ∈ Rn , if and only if their
corresponding matrices G and G′ satisfy G = S ⊤ G′ S. Moreover, this occurs
for some S if and only if G and G′ have the same rank and the same number
of nonnegative eigenvalues—this is Sylvester [290]’s ‘law of inertia for quadratic
forms’, cf. Gantmacher [149], Vol. 1, Chapter X, §2. We shall now be concerned
with matrices that have nonnegative eigenvalues only.
Lemma 2.9.1 Let G be a real symmetric n × n-matrix. Equivalent are:
(i) For all x ∈ Rn , x⊤ Gx ≥ 0.
(ii) All eigenvalues of G are nonnegative.
(iii) G can be written as G = H ⊤ H, with H an m × n matrix, where m is the

rank of G.
Proof. There is an orthogonal matrix Q and a diagonal matrix D whose

nonzero entries are the eigenvalues of G such that G = Q⊤ DQ. If (ii) holds,
then x⊤ Gx = (Qx)⊤ D(Qx) ≥ 0 implies (i). Conversely, (ii) follows from (i) by
choosing x to be an eigenvector. If G = H ⊤ H then x⊤ Gx = ||Hx||2 ≥ 0, so
(iii) implies (i). Finally, let E = D1/2 be the diagonal matrix that squares to
D, and let F be the m × n matrix obtained from E by dropping the zero rows.
Then G = Q⊤ E ⊤ EQ = Q⊤ F ⊤ F Q = H ⊤ H, so that (ii) implies (iii).
A symmetric n × n-matrix G satisfying (i) or (ii) is called positive semidefinite.
It is called positive definite when x⊤ Gx = 0 implies x = 0, or, equivalently,
when all its eigenvalues are positive. For any collection X of vectors of Rm , we
define its Gram matrix as the square matrix G indexed by X whose (x, y)-entry
Gxy is the inner product (x, y) = x⊤ y. This matrix is always positive semidefi-
nite, and it is definite if and only if the vectors in X are linearly independent.
(Indeed, if n = |X|, and we use H to denote the m × n-matrix whose columns
are the vectors of X, then G = H ⊤ H, and x⊤ Gx = ||Hx||2 ≥ 0.)
Lemma 2.9.2 Let N be a real m × n matrix. Then the matrices N N ⊤ and

N ⊤ N have the same nonzero eigenvalues (including multiplicities). Moreover,
rk N N ⊤ = rk N ⊤ N = rk N .
Proof. Let θ be a nonzero eigenvalue of N N ⊤ . The map u 7→ N ⊤ u is an

isomorphism from the θ-eigenspace of N N ⊤ onto the θ-eigenspace of N ⊤ N .
Indeed, if N N ⊤ u = θu then N ⊤ N N ⊤ u = θN ⊤ u and N ⊤ u is nonzero for
nonzero u since N N ⊤ u = θu. The final sentence follows since rk N ⊤ N ≤ rk N ,
but if N ⊤ N x = 0 then ||N x||2 = x⊤ N ⊤ N x = 0, so that N x = 0.
2.10 Diagonally dominant matrices

A diagonally dominant
P matrix is a complex matrix B with the property that
we have |bii | ≥ j6=i |bij | for all i. When all these inequalities are strict, the
matrix is called strictly diagonally dominant.
Lemma 2.10.1 (i) A strictly diagonally dominant complex matrix is nonsin-

gular.
(ii) A symmetric diagonally dominant real matrix with nonnegative diagonal
entries is positive semidefinite.
(iii) Let B be a symmetric real matrix with nonnegative row sums and non-
positive off-diagonal entries. Define a graph Γ on the index set of the rows of
B, where two distinct indices i, j are adjacent when bi,j 6= 0. The multiplicity of
the eigenvalue 0 of B equals the number of connected components C of Γ such
that all rows i ∈ C have zero row sum.
Proof. Let B = (bij ) be diagonally dominant, and let u be an eigenvector,

Pwith Bu = bu. Let |ui | be maximal among the |uj |. Then (bii − b)ui =
say,
− j6=i bij uj . In all cases the result follows by comparing the absolute values
of both sides.
In order to prove (i), assume that B is singular, P and that BuP= 0. Take ab-
solute values on both sides. We find |bii |.|ui | ≤ j6=i |bij |.|uj | ≤ j6=i |bij ||ui | <
|bii |.|ui |. Contradiction.
For (ii), assume that B has a negative eigenvalue b. Then (bii − b).|ui | ≤
|bii |.|ui |. Contradiction.
For (iii), take
P b = 0 again,Pand see how equality could hold everywhere
in bii .|ui | ≤ j6=i |bij |.|uj | ≤ j6=i |bij ||ui | ≤ bii .|ui |. We see that u must
be constant on the connected components of Γ, and zero where row sums are
nonzero.
2.10.1 Geršgorin circles

The above can be greatly generalized. Let B(c, r) = {z ∈ C | |z − c| ≤ r} be
the closed ball in C with center c and radius r.
Proposition 2.10.2 Let AP= (aij ) be a complex matrix of order n, and λ an

eigenvalue of A. Put ri = j6=i |aS
ij |. Then for some i we have λ ∈ B(aii , ri ).
If C is a connected component of i B(aii , ri ) that contains m of the aii , then
C contains m eigenvalues of A.
P
Proof. If Au = λu, then (λ − aii )ui =P j6=i aij uj . Let i be an index for which
|ui | is maximal. Then |λ − aii |.|ui | ≤ j6=i |aij |.|ui | so that λ ∈ B(aii , ri ). For
the second part, use that the eigenvalues are continuous functions of the matrix
elements. Let A(ε) be the matrix with the same diagonal as A and with off-
diagonal entries εaij , so that A = A(1). Then A(0) S has eigenvalues aii , and for
0 ≤ ε ≤ 1 the matrix A(ε) has eigenvalues inside i B(aii , ri ).
This result is due to Geršgorin [151]. A book-length treatment was given by
Varga [300].
2.11. PROJECTIONS 37
2.11 Projections

Q N
Lemma 2.11.1 Let P = be a real symmetric matrix of order n
N⊤ R
with two eigenvalues a and b, partitioned with square Q and R. Let Q have h
eigenvalues θj distinct from a and b. Then R has h eigenvalues a+b−θj distinct
from a and b, and h = mP (a) − mQ (a) − mR (a) = mP (b) − mQ (b) − mR (b),
where mM (η) denotes the multiplicity of the eigenvalue η of M .
Proof. W.l.o.g. a = 1 and b = 0 so that P is a projection and P 2 = P . Now

if Qu = θu then Rv = (1 − θ)v for v = N ⊤ u and N N ⊤ u = θ(1 − θ)u, so that the
eigenvalues of Q and R different from 0 and 1 correspond 1-1. The rest follows
by taking traces: 0 = tr P − tr Q − tr R = mP (1) − mQ (1) − mR (1) − h.
2.12 Exercises
Exercise 1 Consider a symmetric n × n matrix A with eigenvalues θ1 , . . . , θn .
Suppose A has an equitable partition {X1 , . . . , Xm }, where all classes have equal
size. Let S and B be the characteristic matrix and the quotient matrix of this
partition, respectively. Assume that θ1 , . . . , θn are ordered such that θ1 , . . . , θm
are the eigenvalues of B. Prove that A and SS ⊤ commute and give an expression
for the eigenvalues of A + αSS ⊤ for α ∈ R.
Exercise 2 Let B denote the quotient matrix of a symmetric matrix A whose
rows and columns are partitioned according to a partitioning {X1 , . . . , Xm }.
(i) Give an example, where the eigenvalues of B are a sub(multi)set of the

eigenvalues of A, whilst the partition is not equitable.
(ii) Give an example where the partition is equitable, whilst the interlacing is
not tight.
Exercise 3 Let Γ be an undirected graph with smallest eigenvalue −1. Show

that Γ is the disjoint union of complete graphs.
Chapter 3
Graph spectrum (II)
In this chapter we apply the linear algebra from the previous chapter to graph
spectra.
3.1 The largest eigenvalue

The largest eigenvalue of a graph is also known as its spectral radius or index.
The basic information about the largest eigenvalue of a (possibly directed)
graph is provided by Perron-Frobenius theory.
Proposition 3.1.1 Each graph Γ has a real eigenvalue θ0 with nonnegative real
corresponding eigenvector, and such that for each eigenvalue θ we have |θ| ≤ θ0 .
The value θ0 (Γ) does not increase when vertices or edges are removed from Γ.
Assume that Γ is strongly connected. Then
(i) θ0 has multiplicity 1.
(ii) If Γ is primitive (strongly connected, and such that not all cycles have
a length that is a multiple of some integer d > 1), then |θ| < θ0 for all
eigenvalues θ different from θ0 .
(iii) The value θ0 (Γ) decreases when vertices or edges are removed from Γ.
Now let Γ be undirected. By Perron-Frobenius theory and interlacing we

find an upper and lower bound for the largest eigenvalue of a connected graph.
(Note that A is irreducible if and only if Γ is connected.)
Proposition 3.1.2 Let Γ be a connected graph with largest eigenvalue θ1 . If Γ

is regular of valency k, then θ1 = k. Otherwise, we have kmin < k̄ < θ1 < kmax
where kmin , kmax and k̄ are the minimum, maximum and average degree.
Proof. Let 1 be the vector with all entries equal to 1. Then A1 ≤ kmax 1, and
by Theorem 2.2.1(iv) we have θ1 ≤ kmax with equality if and only if A1 = θ1 1,
that is, if and only if Γ is regular of degree θ1 .
Now consider the partition of the vertex set consisting of a single part. By
Corollary 2.5.4 we have k̄ ≤ θ1 with equality if and only if Γ is regular.
39
40 CHAPTER 3. GRAPH SPECTRUM (II)
For not necessarily connected graphs, we have k̄ ≤ θ1 ≤ kmax , and k̄ = θ1 if

and only if Γ is regular. If θ1 = kmax then we only know that Γ has a regular
component with this valency, but Γ need not be regular itself.
As was noted already in Proposition 3.1.1, the largest eigenvalue of a con-
nected graph decreases strictly when an edge is removed.
3.1.1 Graphs with largest eigenvalue at most 2

As an example of the application of Theorem 2.2.1 we can mention:
Theorem 3.1.3 (Smith [285], cf. Lemmens & Seidel [221]). The only con-
nected graphs having largest eigenvalue 2 are the following graphs (the number
of vertices is one more than the index given).
Ân (n ≥ 2) •1
1
• •2
• • • • Ê6 • • • • •
1 1 1 1 1 2 3 2 1
•2
D̂n (n ≥ 4) Ê7 • • • • • • •
1 2 3 4 3 2 1
•1 •1 •3
• • • • • • Ê8 • • • • • • • •
1 2 2 2 2 1 2 4 6 5 4 3 2 1
For each graph, the corresponding eigenvector is indicated by the integers at the
vertices. Moreover, each connected graph with largest eigenvalue less than 2 is
a subgraph of one of the above graphs, i.e., one of the graphs An = Pn , the path
with n vertices (n ≥ 1), or
E6 • • • • •
Dn (n ≥ 4) •
• E7 • • • • • •
• • • • • • •
E8 • • • • • • •
Finally, each connected graph with largest eigenvalue more than 2 contains
one of Ân , D̂n , Ê6 , Ê7 , Ê8 as a subgraph.
Proof. The vectors indicated are eigenvectors for the eigenvalue 2. Therefore,
Ân , D̂n and Êm (m = 6, 7, 8) have largest eigenvalue 2. Any graph containing
one of these as an induced proper subgraph has an eigenvalue larger than 2. So,
if Γ has largest eigenvalue at most 2 and is not one of Ân or D̂n , then Γ is a
3.1. THE LARGEST EIGENVALUE 41
tree without vertices of degree at least 4 and with at most one vertex of degree
three, and the result easily follows.
These graphs occur as the Dynkin diagrams and extended Dynkin diagrams of
finite Coxeter groups, cf. [35, 48, 203]. Let us give their eigenvalues:
The eigenvalues of An are 2 cos iπ/(n + 1) (i = 1, 2, ..., n).
The eigenvalues of Dn are 0 and 2 cos iπ/(2n − 2) (i = 1, 3, 5, ..., 2n − 3).
The eigenvalues of E6 are 2 cos iπ/12 (i = 1, 4, 5, 7, 8, 11).
The eigenvalues of E7 are 2 cos iπ/18 (i = 1, 5, 7, 9, 11, 13, 17).
The eigenvalues of E8 are 2 cos iπ/30 (i = 1, 7, 11, 13, 17, 19, 23, 29).
(Indeed, these eigenvalues are 2 cos(di − 1)π/h (1 ≤ i ≤ n) where h is the
Coxeter number, and the di are the degrees, cf. [48, pp. 84, 308]. Note that in
all cases the largest eigenvalue is 2 cos π/h.)
The eigenvalues of D̂n are 2, 0, 0, −2 and 2 cos iπ/(n − 2) (i = 1, ..., n − 3).
The eigenvalues of √ 1, 0, −1, −1, −2.
Ê6 are 2, 1, √
The eigenvalues of Ê7 are 2, 2, 1, 0, 0, −1, − 2, −2.
The eigenvalues of Ê8 are 2, τ , 1, τ −1 , 0, −τ −1 , −1, −τ , −2.
Remark It is possible to p go a√little bit further, and find all graphs with
largest eigenvalue at most 2 + 5 ≈ 2.05817, cf. √ Brouwer & Neumaier [59].
For the graphs with largest eigenvalue at most 32 2 ≈ 2.12132, see Woo &
Neumaier [308] and Cioabă, van Dam, Koolen & Lee [91].
3.1.2 Subdividing an edge

Let Γ be a graph on n vertices, and consider the graph Γ′ on n + 1 vertices
obtained from Γ by subdividing an edge e (that is, by replacing the edge e = xy
by the two edges xz and zy where z is a new vertex). The result below relates
the largest eigenvalue of Γ and Γ′ .
We say that e lies on an endpath if Γ \ e (the graph on n vertices obtained
by removing the edge e from Γ) is disconnected, and one of its connected com-
ponents is a path.
Proposition 3.1.4 (Hoffman-Smith [197]) Let Γ be a connected graph, and let
the graph Γ′ be obtained from Γ by subdividing an edge e. Let Γ and Γ′ have
largest eigenvalues λ and λ′ , respectively. Then if e lies on an endpath, we have
λ′ > λ, and otherwise λ′ ≤ λ, with equality only when both equal 2.
Proof. If e lies on an endpath, then Γ is obtained from Γ′ by removing a
leaf vertex, and λ < λ′ follows by Proposition 3.1.1. Suppose e is not on an
endpath. By Theorem 3.1.3, λ ≥ 2. Let A and A′ be the adjacency matrices of
Γ and Γ′ , so that Au = λu for some vector u > 0. We use Theorem 2.2.1 (iv)
and conclude λ′ ≤ λ from the existence of a nonzero vector v with v ≥ 0 and
A′ v ≤ λv. Such a vector v can be constructed as follows. If z is the new point
on the edge e = xy, then we can take vp = up for p 6= z, and vp = min(ux , uy ),
provided that λvp ≥ ux + uy . Suppose not. W.l.o.g.,
P assume ux ≤ uy , so that
λux < ux +uy , and hence ux < uy . We have 0 ≤ p ∼ x,p6=y up = λux −uy < ux .
If x has degree 2 in Γ, say x ∼ p, y, then replace e = xy by e = px to decrease
the values of u on the end points of e — this does not change Γ′ . If x has
degree m > 2, then construct v by vx = λux − uy and vz = ux and vp = up
for p 6= Px, z. We have to check that λvx ≥ vx + ux , but this follows from
λvx = λ p ∼ x,p6=y up ≥ (m − 1)ux ≥ 2ux > vx + ux .
3.1.3 The Kelmans operation

As we saw, adding edges causes the largest eigenvalue to increase. The operation
described below (due to Kelmans [215]) only moves edges, but also increases θ1 .
Given a graph Γ and two specified vertices u, v construct a new graph Γ′ by
replacing the edge vx by a new edge ux for all x such that v ∼ x 6∼ u. The new
graph Γ′ obtained in this way has the same number of vertices and edges as the
old graph, and all vertices different from u, v retain their valency. The vertices
u, v are adjacent in Γ′ if and only if they are adjacent in Γ. An isomorphic graph
is obtained if the rôles of u and v are interchanged: if N (u) and N (v) are the
sets of neighbours of u, v distinct from u and v, then in the resulting graph the
corresponding sets are N (u) ∪ N (v) and N (u) ∩ N (v).
If Γ denotes the complementary graph of Γ then also Γ′ is obtained by a
Kelmans operation from Γ.
Proposition 3.1.5 (Csikvári [101]) Let Γ be a graph, and let Γ′ be obtained

from Γ by a Kelmans operation. Then θ1 (Γ) ≤ θ1 (Γ′ ). (And hence also θ1 (Γ) ≤
θ1 (Γ′ ).)
Proof. Let A and A′ be the adjacency matrices of Γ and Γ′ , and let Ax = θ1 x

where x ≥ 0, x⊤ xP = 1. W.l.o.g., let xu ≥ xv . Then θ1 (Γ′ ) ≥ x⊤ A′ x =
x⊤ Ax + 2(xu − xv ) w∈N (v)\N (u) xw ≥ θ1 (Γ).
√
Csikvári continues and uses this to show that θ1 (Γ) + θ1 (Γ) ≤ 21 (1 + 3)n.
Earlier, Brualdi & Hoffman [63] had observed that a graph with maximal spec-
tral radius ρ among the graphs with a given number of vertices and edges has
a vertex ordering such that if x ∼ y and z ≤ x, w ≤ y, z 6= w, then z ∼ w.
Rowlinson [263] calls the adjacency matrices of these graphs (ordered this way)
stepwise and proves that the maximal value of ρ among the graphs on n vertices
and e edges
Km + (n − m)K1 , where m is minimal such
is obtained by taking
that m 2 ≥ e, and removing m
2 − e edges on a single vertex.
It follows from the above proposition that a graph with maximal θ1 (Γ)+θ1 (Γ)
has stepwise matrix. It is conjectured that in fact θ1 (Γ) + θ1 (Γ) < 34 n − 1.
3.2 Interlacing
By Perron-Frobenius theory, the largest eigenvalue of a connected graph goes
down when one removes an edge or a vertex. Interlacing also gives information
about what happens with the other eigenvalues.
The pictures for A and L differ. The eigenvalues for the adjacency matrix
A show nice interlacing behavior when one removes a vertex, but not when an
edge is removed. (Cf. §1.9.) The Laplace eigenvalues behave well in both cases.
For A an eigenvalue can go both up or down when an edge is removed. For L
it cannot increase.
Proposition 3.2.1 (i) Let Γ be a graph and ∆ an induced subgraph. Then the
eigenvalues of ∆ interlace those of Γ,
(ii) Let Γ be a graph and let ∆ be a subgraph, not necessarily induced, on m
vertices. Then the i-th largest Laplace eigenvalue of ∆ is not larger than the i-th
largest Laplace eigenvalue of Γ (1 ≤ i ≤ m), and the i-th largest signless Laplace
3.3. REGULAR GRAPHS 43
eigenvalue of ∆ is not larger than the i-th largest signless Laplace eigenvalue of
Γ (1 ≤ i ≤ m).
Proof. Part (i) is immediate from Corollary 2.5.2. For part (ii), recall that we
have L = N N ⊤ when N is the directed point-edge incidence matrix obtained
by orienting the edges of Γ arbitrarily, and that N N ⊤ and N ⊤ N have the
same nonzero eigenvalues. Removing an edge from Γ corresponds to removing
a column from N , and leads to a principal submatrix of N ⊤ N , and interlacing
holds. Removing an isolated vertex from Γ corresponds to removing a Laplace
eigenvalue 0. The same proof applies to the signless Laplace matrix.
3.3 Regular graphs

It is possible to see from the spectrum whether a graph is regular:
Proposition 3.3.1 Let Γ be a graph with eigenvalues k = θ1 ≥ θ2 ≥ ... ≥ θv .

Equivalent are:
(i) Γ is regular (of degree k),
(ii) AJ = kJ,
P 2
(iii) θi = kv.
Proof. We have P seen that (i) and (ii) are equivalent. Also, if Γ is regular
of degree
P 2 k, then θi2 = tr A2 = kv. Conversely, if (iii) holds, then k̄ =
−1
v θi = θ1 and, by Proposition 3.1.2, Γ is regular.
As we saw above in §1.3.7, it is also possible to see from the spectrum whether
a graph is regular and connected. However, for nonregular graphs it is not
possible to see from the spectrum whether it is connected.
A very useful characterization of regular connected graphs was given by
Hoffman [194]:
Proposition 3.3.2 The graph Γ is regular and connected if and only if there
exists a polynomial p such that J = p(A).
Proof. If J = p(A), then J commutes with A and hence Γ is regular (and

clearly also connected). Conversely, let Γ be connected and regular. Choose
a basis such that the commuting matrices A and J become diagonal. Then
J becomes
Qv diag(v, 0, ..., 0) and A becomes diag(k, θ2 , ..., θv ). Hence, if we put
f (x) = i=2 (x − θi ), then J = vf (A)/f (k), so that p(x) = vf (x)/f (k) satisfies
the requirements.
3.4 Bipartite graphs

Among the connected graphs Γ, those with imprimitive A are precisely the
bipartite graphs (and for these, A has period 2). Consequently we find from
Theorem 2.2.1(iii):
Proposition 3.4.1 (i) A graph Γ is bipartite if and only if for each eigenvalue
θ of Γ, −θ is also an eigenvalue, with the same multiplicity.
(ii) If Γ is connected with largest eigenvalue θ1 , then Γ is bipartite if and
only if −θ1 is an eigenvalue of Γ.
Proof. For connected graphs all is clear from the the Perron-Frobenius theo-
rem. That gives (ii) and (by taking unions) the ‘only if’ part of (i). For the ‘if’
part of (i), let θ1 be the spectral radius of Γ. Then some connected component
of Γ has eigenvalues θ1 and −θ1 , and hence is bipartite. Removing its contri-
bution to the spectrum of Γ we see by induction on the number of components
that all components are bipartite.
3.5 Classification of integral cubic graphs

As an application of Proposition 3.3.2, let us classify the cubic graphs (graphs,
that are regular of valency 3) with integral spectrum. The result is due to
Bussemaker & Cvetković [68]. See also Schwenk [268]. There are 13 examples.
case v spectrum description

i 6 ±3, 04 K3,3
ii 8 ±3, (±1)3 23
iii 10 ±3, ±2, (±1)2 , 02 ∗
K2,3 ⊗ K2
iv 12 ±3, (±2)2 , ±1, 04 C 6 × K2
v 20 ±3, (±2)4 , (±1)5 Π ⊗ K2
vi 20 ±3, (±2)4 , (±1)5 T ∗ ⊗ K2
vii 24 ±3, (±2)6 , (±1)3 , 04 Σ ⊗ K2
viii 30 ±3, (±2)9 , 010 GQ(2, 2)
ix 4 3, (−1)3 K4
x 6 3, 1, 02 , (−2)2 K3 × K2
xi 10 3, 15 , (−2)4 Π
xii 10 3, 2, 13 , (−1)2 , (−2)3 (Π ⊗ K2 )/σ
xiii 12 3, 23 , 02 , (−1)3 , (−2)3 Σ
A quotient of the hexagonal grid

Let us describe a graph that comes up in the classification. Take a tetrahedron
and cut off each corner. Our graph Σ is the 1-skeleton of the resulting polytope,
or, equivalently, the result of replacing each vertex of K4 by a triangle (a Y − ∆
operation). It can also be described as the line graph of the graph obtained
from K4 by subdividing each edge. The bipartite double Σ ⊗ K2 of Σ is more
beautiful (for example, its group is a factor 6 larger than that of Σ), and can
be described as the quotient Λ/6Λ of the hexagonal grid Λ = ha + bω | a, b ∈
Z, a + b = 0, 1(mod 3)i in the complex plane, with ω 2 + ω + 1 = 0. Now Σ is
found e.g. as Λ/h3a + 6bω | a, b ∈ Zi.
Cubic graphs with loops

For a graph Γ where all vertices have degree 2 or 3, let Γ∗ be the cubic graph
(with loops) obtained by adding a loop at each vertex of degree 2. Note that the
3.5. CLASSIFICATION OF INTEGRAL CUBIC GRAPHS 45
sum of the eigenvalues of Γ∗ , the trace of its adjacency matrix, is the number
of loops.
∗
The graph K2,3 has spectrum 3, 1, 1, 0, −2.
Let T be the graph on the singletons and pairs in a 4-set, where adjacency
is inclusion. Then T ∗ has spectrum 31 , 23 , 12 , (−1)3 , (−2)1 .
The classification
Let Π be the Petersen graph and Σ, T the graphs described above.
We split the result into two propositions, one for the bipartite and one for
the nonbipartite case.
Proposition 3.5.1 Let Γ be a connected bipartite cubic graph such that all of
its eigenvalues are integral. Then Γ is one of 8 possible graphs, namely (i) K3,3 ,
(ii) 23 , (iii) K2,3
∗
⊗ K2 , (iv) C6 × K2 , (v) the Desargues graph (that is, the
bipartite double Π ⊗ K2 of the Petersen graph Π), (vi) T ∗ (cospectral with the
previous), (vii) the bipartite double of Σ, (viii) the point-line incidence graph
of the generalized quadrangle of order 2 (that is, the unique 3-regular bipartite
graph with diameter 4 and girth 8, also known as Tutte’s 8-cage).
Proof. Let Γ have spectrum (±3)1 (±2)a (±1)b 02c (with multiplicities written
as exponents).
The total number of vertices is v = 2 + 2a + 2b + 2c. The total number of
edges is 23 v = 21 trA2 = 9 + 4a + b (so that 2b + 3c = 6 + a). The total number
of quadrangles is q = 9 − a − b, as one finds by computing trA4 = 15v + 8q =
2(81+16a+b). The total number of hexagons is h = 10+2b−2c, found similarly
by computing trA6 = 87v + 96q + 12h = 2(729 + 64a + b).
Somewhat more detailed, let qu be the number of quadrangles on the vertex
u, and quv the number of quadrangles on the edge uv, and similarly for h. Let
uv be an edge. Then Auv = 1 and (A3 )uv = 5 + quv and (A5 )uv = 29 + 2qu +
2qv + 6quv + huv .
The Hoffman polynomial A(A+3I)(A2 −I)(A2 −4I) defines a rank 1 matrix
with eigenvalue 720, so that A(A+3I)(A2 −I)(A2 −4I) = 720 v J and in particular
v|240 since for an edge xy the xy entry of 720 v J must be divisible by 3. This
leaves for (a, b, c, v) the possibilities a) (0, 0, 2, 6), b) (0, 3, 0, 8), c) (1, 2, 1, 10),
d) (2, 1, 2, 12), e) (3, 3, 1, 16), f) (4, 5, 0, 20), g) (5, 1, 3, 20), h) (6, 3, 2, 24), i)
(9, 0, 5, 30).
In case a) we have K3,3 , case (i) of the theorem.
In case b) we have the cube 23 , case (ii).
In case c) we have a graph of which the bipartite complement has spectrum
22 12 02 (−1)2 (−2)2 hence is the disjoint union of a 4-cycle and a 6-cycle, case
(iii).
In case d) we have q = 6 and h = 8. Let uv be an edge, and evaluate
A(A+3I)(A2 −I)(A2 −4I) = 60J at the uv position to find (A5 −5A3 +4A)uv =
20 and 2qu + 2qv + quv + huv = 12. It follows that uv cannot lie in 3 or more
quadrangles. Suppose u lies in (at least) 3 quadrangles. Then for each neighbor
x of u we have 2qx + hux = 4 so that qx = 2 and hu = 0. The mod 2 sum of
two quadrangles on u is not a hexagon, and it follows that we have a K2,3 on
points u, w, x, y, z (with u, w adjacent to x, y, z). The six quadrangles visible in
the K2,3 on u, w, x, y, z contribute 6 + 4 + 2 + 0 to 2qu + 2qx + qux + hux = 12,
and it follows that there are no further quadrangles or hexagons on these points.
So the three further neighbors p, q, r of x, y, z are distinct and have no common

neighbors, impossible since v = 12. So, no vertex is in 3 or more quadrangles,
and hence every vertex u is in precisely 2 quadrangles. These two quadrangles
have an edge uu′ in common, and we find an involution interchanging each u
and u′ , and preserving the graph. It follows that we either have C6 × K2 (and
this has the desired spectrum, it is case (iv)), or a twisted version, but that has
only 6 hexagons.
In case e) we have v = 16 vertices. For any vertex x, Hoffman’s polynomial
yields (A6 )xx − 5(A4 )xx + 4(A2 )xx = 45. On the other hand, (A2i )xx is odd for
each i, since each walk of length 2i from x to x can be paired with the reverse
walk, so that the parity of (A2i )xx is that of the number of self-reverse walks
x...zwz...x which is 3i . Contradiction.
In case f) we have v = 20, q = 0, h = 20. Since c = 0 we can omit the factor
A from Hoffman’s polynomial and find (A + 3I)(A2 − I)(A2 − 4I) = 12J. If u, w
have even distance, then (A4 −5A2 +4I)uw = 4. In particular, if d(u, w) = 2 then
9 = (A4 )uw = 7 + huw so that huw = 2: each 2-path uvw lies in two hexagons.
If no 3-path uvwx lies in two hexagons then the graph is distance-regular with
intersection array {3, 2, 2, 1, 1; 1, 1, 2, 2, 3} (cf. Chapter 11) and hence is the
Desargues graph. This is case (v) of the theorem. Now assume that the 3-path
uvwx lies in two hexagons, so that there are three paths u ∼ vi ∼ wi ∼ x
(i = 1, 2, 3). The vi and wi need one more neighbor, say vi ∼ yi and wi ∼ zi
(i = 1, 2, 3). The vertices yi are distinct since there are no quadrangles, and
similarly the zi are distinct. The vertices yi and zj are nonadjacent, otherwise
the would be a quadrangle (if i = j) or uvj wj would be in three hexagons (if
i 6= j). Remain 6 more vertices, 3 each adjacent to two vertices yi and 3 each
adjacent to two vertices zi . Call them si and ti , where si ∼ yj and ti ∼ zj
whenever i 6= j. The final part is a matching between the si and the ti . Now
the 2-path vi wi zi is in two hexagons, and these must be of the form tj zi wi vi yi sk
with j 6= i 6= k, and necessarily j = k, that is, the graph is uniquely determined.
This is case (vi) of the theorem.
In case g) we have v = 20, q = 3, h = 6. For an edge uv we have (A5 −
3
5A + 4A)uv = 12, so that 2qu + 2qv + quv + huv = 4. But that means that the
edge uv cannot be in a quadrangle, contradiction.
In case h) we have v = 24, q = 0, h = 12. For an edge uv we have (A5 −5A3 +
4A)uv = 10, so that huv = 2. It follows that each vertex is in 3 hexagons, and
each 2-path vuw is in a unique hexagon. Now one straightforwardly constructs
the unique cubic bipartite graph on 24 vertices without quadrangles and such
that each 2-path is in a unique hexagon. Starting from a vertex u, call its
neighbors vi (i = 1, 2, 3), call the six vertices at distance two wij (i, j = 1, 2, 3
and i 6= j), and let xi (i = 1, 2, 3) be the three vertices opposite u in a hexagon
on u, so that the three hexagons on u are uvi wij xk wji vj (with distinct i, j, k).
Let the third neighbor of wij be yij , and let the third neighbor of xk be zk .
Necessarily zk ∼ ykj . Now each vertex yij still needs a neighbor and there are
two more vertices, say s ∼ y12 , y23 , y31 and t ∼ y13 , y21 , y32 . This is case (vii).
In case i) we have v = 30, q = h = 0 and we have Tutte’s 8-cage. This is
case (viii).
Proposition 3.5.2 Let Γ be a connected nonbipartite cubic graph such that all
of its eigenvalues are integral. Then Γ is one of 5 possible graphs, namely (ix)
K4 , (x) K3 × K2 , (xi) the Petersen graph, (xii) the graph on 10 vertices defined
3.6. THE LARGEST LAPLACE EIGENVALUE 47
by i ∼ (i + 1) (mod 10), 0 ∼ 5, 1 ∼ 3, 2 ∼ 6, 4 ∼ 8, 7 ∼ 9 (or, equivalently,

the graph obtained from K3,3 by replacing two nonadjacent vertices by a triangle
with a Y − ∆ operation), (xiii) Σ.
Proof. Consider Γ ⊗ K2 . It is cubic and has integral eigenvalues, hence is one

of the 8 graphs ∆ found in the previous proposition. There is an involution σ
of ∆ = Γ ⊗ K2 without fixed edges, that interchanges the two vertices x′ and
x′′ for each vertex x of Γ. Now Γ can be retrieved as ∆/σ.
In cases (i), (iii), (viii) the graph Γ would be cubic on an odd number of
vertices, impossible.
In case (ii), σ must interchange antipodes, and the quotient 23 /σ is the
complete graph K4 . This is case (ix).
In case (iv), C6 × K2 , σ must interchange antipodes in the same copy of C6 ,
and the quotient is K3 × K2 . This is case (x).
In case (v), Π ⊗ K2 , we get the Petersen graph for a σ that interchanges
antipodal vertices. This is case (xi). The group is Sym(5).2 and has two
conjugacy classes of suitable involutions σ. The second one interchanges x′
with (12)x′′ , and its quotient is obtained from Π by replacing the hexagon
13 ∼ 24 ∼ 15 ∼ 23 ∼ 14 ∼ 25 ∼ 13 by the two triangles 13, 14, 15 and
23, 24, 25. This is case (xii).
In case (vi) there is no suitable σ. (An automorphism σ must interchange
the two vertices u, x found in the previous proof, since this is the only pair of
vertices joined by three 3-paths. But any shortest ux-path is mapped by σ into
a different xu-path (since the path has odd length, and σ cannot preserve the
middle edge) so that the number of such paths, which is 3, must be even.)
In case (vii) we get Σ. This is case (xiii). (The group of Σ ⊗ K2 has order
144, six times the order of the group Sym(4) of Σ, and all possible choices of σ
are equivalent.)
3.6 The largest Laplace eigenvalue

If µ1 ≤ . . . ≤ µn are the Laplace eigenvalues of a simple graph Γ, then 0 ≤
n − µn ≤ . . . ≤ n − µ2 are the the Laplace eigenvalues of the complement of Γ
(see §1.3.2). Therefore µn ≤ n with equality if and only if the complement of Γ
is disconnected. If Γ is regular with valency k we know (by Proposition 3.4.1)
that µn ≤ 2k, with equality if and only if Γ is bipartite. More generally:
Proposition 3.6.1 Let Γ be a graph with adjacency matrix A (with eigenvalues

θ1 ≥ . . . ≥ θn ), Laplacian L (with eigenvalues µ1 ≤ . . . ≤ µn ), and signless
Laplacian Q (with eigenvalues ρ1 ≥ . . . ≥ ρn ). Then
(i) (Zhang & Luo [313])

µn ≤ ρ1 .
If Γ is connected, then equality holds if and only if Γ is bipartite.
(ii) Let dx be the degree of the vertex x. If Γ has at least one edge then
ρ1 ≤ max (dx + dy ).
x∼y
Equality holds if and only if Γ is regular or bipartite semiregular.

(iii) (Yan [309])

2θi ≤ ρi (1 ≤ i ≤ n).
Proof. (i) Apply Theorem 2.2.1 (vi).

(ii) Using Proposition 3.1.2 to bound the largest eigenvalue θ1 (L(Γ)) of L(Γ)
by its maximum degree maxx ∼ y (dx + dy − 2) we find ρ1 = θ1 (L(Γ)) + 2 ≤
maxx ∼ y (dx + dy ), with equality if and only if L(Γ) is regular so that Γ is
regular or bipartite semiregular.
(iii) Since Q = L + 2A and L is positive semidefinite, this follows from the
Courant-Weyl inequalities (Theorem 2.8.1 (iii)).
Corollary 3.6.2 ([8]) Let Γ be a graph on n vertices with at least one edge.
Then
µn ≤ max (dx + dy ).
x∼y
If Γ is connected, then equality holds if and only if Γ is bipartite regular or

semiregular.
For bipartite graphs, L and Q have the same spectrum (see Proposition 1.3.10).
It follows by Perron-Frobenius that the largest Laplace eigenvalue of a connected
bipartite graph decreases strictly when an edge is removed.
Interlacing provides a lower bound for µn :
Proposition 3.6.3 ([166]) Let Γ be a graph on n vertices with at least one edge,
and let dx be the degree of the vertex x. Then
µn ≥ 1 + max dx .
x
If Γ is connected, then equality holds if and only if maxx dx = n − 1.
Proof. If Γ has a vertex of degree d, then it has a subgraph K1,d (not nec-
essarily induced), and µn ≥ d + 1. If equality holds, then Γ does not have a
strictly larger bipartite subgraph. If Γ is moreover connected, then d = n − 1.

Deriving bounds on µn has become an industry—there are many papers, cf. [36,
124, 170, 223, 224, 231, 246, 312].
3.7 Laplace eigenvalues and degrees
The Schur inequality (Theorem 2.6.1) immediately yields an inequality between

the sum of the largest m Laplace eigenvalues and the sum of the largest m
vertex degrees. Grone [165] gave a slightly stronger result.
Proposition 3.7.1 If Γ is connected, with Laplace eigenvalues ν1 ≥ ν2 ≥ ... ≥

νn = 0 and Pm d1 ≥ d2 ≥ ... ≥ dn > 0, then for 1 ≤ m ≤ n − 1 we
Pmvertex degrees
have 1 + i=1 di ≤ i=1 νi .
3.7. LAPLACE EIGENVALUES AND DEGREES 49
Proof. Let xi have degree di and put Z = {x1 , . . . , xm }. Let N (Z) be the
set of vertices outside Z with a neighbor in Z. Instead of assuming that Γ is
connected we just
P use that N (Z) is nonempty. P If we delete the vertices outside
m
Z ∪N (Z) then z∈Z dz does not change, while i=1 νi does not increase, so we
may assume X = Z ∪ N (Z). Let R be the quotient matrix of L for the partition
{{z} | z ∈ Z} ∪ {N (Z)} of X, and let λ1 ≥ . . . ≥ λm+1 be the eigenvalues of R.
The matrix R has row sums 0, so λm+1 = 0. By interlacing (Corollary 2.5.4)
Pm Pm Pm+1 P
we have i=1 νi ≥ i=1 λi = i=1 λi = tr R = z∈Z dz + Rm+1,m+1 and the
desired result follows since Rm+1,m+1 ≥ 1.
Second proof. We prove the following stronger statement:
For any graph Γ (notPnecessarilyPconnected) and any subset Z of the vertex
m
set X of Γ one has h + z∈Z dz ≤ i=1 νi , where dz denotes the degree of the
vertex z in Γ, and m = |Z|, and h is the number of connected components of
the graph ΓZ induced on Z that are not connected components of Γ.
We may assume that Γ is connected, and that Z and X \ Z are nonempty.
Now h is the number of connected components of ΓZ .
B −C
The partition {Z, X \ Z} of X induces a partition L = . Since
−C ⊤ E
Γ is connected, B is nonsingular by Lemma 2.10.1(iii). All entries of B −1 are
nonnegative. (Write B = n(I−T ) where T ≥ 0, then B −1 = n1 (I+T +T 2 +. . .) ≥
0. If h = 1, then B −1 > 0.)
P 0
Since L is positive semidefinite, we can write L = M M ⊤ , where M =
Q R
is a square matrix. Now B = P P ⊤ and −C = P Q⊤ . The eigenvalues of
M M ⊤ are the same as those of M ⊤ M , and that latter matrix Pm has submatrix
P ⊤ P + QP⊤
Q of order m. By Schur’s inequality we get i=1 νi ≥ tr(P ⊤ P +
Q⊤ Q) = z∈Z dz + tr Q⊤ Q, and it remains to show that tr Q⊤ Q ≥ h.
Now Q⊤ Q = P −1 CC ⊤ P −⊤ , so tr Q⊤ Q = tr B −1 CC ⊤ . We have B = LZ +
D, where LZ is the Laplacian of ΓZ , and D is the diagonal matrix of the row
sums of C. Since CC ⊤ ≥ D and B −1 ≥ 0, we have tr Q⊤ Q ≥ trB −1 D. If
LZ u = 0 then (LZ + D)−1 Du = u. Since LZ has eigenvalue 0 with multiplicity
h, B −1 D has eigenvalue 1 with multiplicity h. Since this matrix is positive
semidefinite (since D1/2 B −1 D1/2 is p.s.d.), its trace is at least h.
A lower bound for the individual νj was conjectured by Guo [171] and proved
in Brouwer & Haemers [54].
Proposition 3.7.2 Let Γ be a graph with Laplace eigenvalues ν1 ≥ ν2 ≥ ... ≥

νn = 0 and with vertex degrees d1 ≥ d2 ≥ ... ≥ dn . Let 1 ≤ m ≤ n. If Γ is not
Km + (n − m)K1 , then νm ≥ dm − m + 2.
We saw the special case m = 1 in Proposition 3.6.3. The cases m = 2 and

m = 3 were proved earlier in [222] and [171].
Examples with equality are given by complete graphs Km with a pending
edges at each vertex (where a >p0), with Laplace spectrum consisting of 0,
1m(a−1) , a + 1 and 21 (m + a + 1 ± (m + a + 1)2 − 4m) with multiplicity m − 1
each, so that νm = a + 1 = dm − m + 2.
Further examples are complete graphs Km with a pending edges attached
at a single vertex. Here n = m + a, the Laplace spectrum consists of m + a,
mm−2 , 1a and 0, so that νm = 1 = dm − m + 2.
Any graph contained in Ka,a and containing K2,a has ν2 = a = d2 , with

equality for m = 2.
Any graph on n vertices with d1 = n − 1 has equality for m = 1.
More generally, whenever one has an eigenvector u and vertices x, y with
ux = uy , then u remains eigenvector, with the same eigenvalue, if we add or
remove an edge between x and y. Many of the above examples can be modified
by adding edges. This leads to many further cases of equality.
3.8 The Grone-Merris Conjecture

3.8.1 Threshold graphs
A threshold graph is a graph obtained from the graph K0 by a sequence of
operations of the form (i) add an isolated vertex, or (ii) take the complement.
Proposition 3.8.1 Let Γ be a threshold graph with Laplace eigenvalues (in
nonincreasing order) ν1 ≥ ν2 ≥ ... ≥ νn = 0. Let dx be the degree of the
vertex x. Then
νj = #{x | dx ≥ j}.
Proof. Induction on the number of construction steps of type (i) or (ii).
Grone & Merris [166] conjectured that this is the extreme case, and that for all
undirected graphs and all t one has
t
X t
X
νj ≤ #{x | dx ≥ j}.
j=1 j=1
For t = 1 this is immediate from ν1 ≤ n. For t = n equality holds. This

conjecture was proved in Hua Bai [13], see §3.8.2 below. There is a generalization
to higher-dimensional simplicial complexes, see §3.9 below.
A variation on the Grone-Merris conjecture is the following.
Conjecture Let Γ be a graph with e edges
Pt and Laplace eigenvalues
ν1 ≥ ν2 ≥
... ≥ νn = 0. Then for each t we have i=1 νi ≤ e + t+1
2 .
It is easy to see (by induction) that this inequality holds for threshold graphs.
In [180] it is proved for trees, and in case t = 2. In [23] it is shown that there
is a t such that the t-th inequality of this conjecture is sharper than the t-th
Grone-Merris inequality if and only if the graph is non-split. In particular, this
conjecture holds for split graphs. It also holds for regular graphs.
3.8.2 Proof of the Grone-Merris Conjecture

Very recently, Hua Bai [13] proved the Grone-Merris Conjecture. We repeat the
statement of the theorem.
Theorem 3.8.2 Let Γ be an undirected graph with Laplace eigenvalues (in non-
increasing order) ν1 ≥ ν2 ≥ ... ≥ νn = 0. Let dx be the degree of the vertex x.
Then for all t, 0 ≤ t ≤ n, we have
t
X t
X
νi ≤ #{x | dx ≥ i}. (3.1)
i=1 i=1
3.8. THE GRONE-MERRIS CONJECTURE 51
The proof is by reducing to the case of a semi-bipartite graph (or split graph),
that is a graph where the vertex set is the disjoint union of a nonempty subset
inducing a clique (complete graph), and a nonempty subset inducing a coclique
(edgeless graph). Then for semi-bipartite graphs a continuity argument proves
the crucial inequality stated in the following lemma.
Lemma 3.8.3 Let Γ be a semi-bipartite graph with clique of size c and Laplace
eigenvalues be ν1 ≥ ν2 ≥ ... ≥ νn = 0. Let δ be the maximum degree among the
vertices
Pc the coclique, so that δ ≤ c. If νc > c or νc = c > δ then we have
in P
c
ν
i=1 i ≤ i=1 #{x | dx ≥ i}.
Proof of Theorem 3.8.2 (assuming Lemma 3.8.3). Consider counterexamples

to (3.1) with minimal possible t.
Step 1 If Γ is such a counterexample with minimal number of edges, and x, y
are vertices in Γ of degree at most t, then they are nonadjacent.
P∼
Indeed, if x y then let Γ′ be the graph
Pt obtained from Γ by removing the edge
t
xy. Then i=1 #{x | d′x ≥ i} + 2 = i=1 #{x | dx ≥ i}. The Laplace matrices
L and L′ of Γ and Γ′ satisfy
Pt L = L P
′
+ H where H has eigenvalues 2, 0n−1 . By
t
Theorem 2.8.2 we have i=1 νi ≤ i=1 νi′ + 2, and since Γ′ has fewer edges
Pt Pt Pt Pt
than Γ we find i=1 νi ≤ i=1 νi′ + 2 ≤ i=1 #{x | d′x ≥ i} + 2 = i=1 #{x |
dx ≥ i}, contradiction.
Step 2 There is a semi-bipartite counterexample Γ for the same t, with clique
size c := #{x | dx ≥ t}.
Indeed, we can form a new graph Γ from the Γ of Step 1 by adding edges
xy
Pt for every pair of nonadjacent vertices x,Py,t both of degree at least t. Now
i=1 #{x | dx ≥ i} does not change, and i=1 νi does not decrease, and the
new graph is semi-bipartite with the stated clique size.
This will be our graph Γ for the rest of the proof.
Step 3 A semi-bipartite graph ∆ of clique size c satisfies νc+1 ≤ c ≤ νc−1 .
Indeed, since ∆ contains the complete graph Kc with Laplace spectrum cc−1 ,
0, we see by the Courant-Weyl inequalities (Theorem 2.8.1 (iii)) that νc−1 ≥ c.
And since ∆ is contained in the complete semi-bipartite graph with clique of
size c and coclique of size n − c and all edges in-between, with Laplace spectrum
nc , cn−c−1 , 0, we have νc+1 ≤ c.
Since t was chosen minimal, we have νt > #{x | dx ≥ t} = c. The previous step
then implies c ≥ t. If c = t then νc > c and Lemma 3.8.3 gives a contradiction.
So c > t.
All vertices in the coclique of Γ have degree at most t − 1 and all vertices in
the clique have degree at least c − 1. So #{x | dx ≥ i} = c for t ≤P i ≤ c − 1.
t
From Step 3 we have νi ≥ νc−1 ≥ c for t ≤ i ≤ c − 1. Since i=1 νi >
Pt Pc−1 Pc−1
i=1 #{x | d x ≥ i} we also have i=1 ν i > i=1 #{x | d x ≥ i}. Now if νc ≥ c
we contradict Lemma 3.8.3 (since #{x | dx ≥ c} ≤ c). So νc < c.
Step 4 The m-th Grone-Merris inequality for a graph Γ is equivalent to the
(n − 1 − m)-th Grone-Merris inequality for its complement Γ (1 ≤ m ≤ n − 1).
Pn(1 ≤ i ≤P
Indeed, Γ has Laplace eigenvalues ν i = n−νn−i n−1) and dual degrees
n
#{x | dx ≥ i} = n − #{x | dx ≥ n − i}, and i=1 νi = i=1 #{x | dx ≥ i}.
In our case Γ is semibipartitePwith cliquePsize n − c, and by the above we have
n−c n−c
ν n−c = n − νc > n − c and i=1 ν i > i=1 #{x | dx ≥ i}. This contradicts
Lemma 3.8.3.
This contradiction completes the proof of the Grone-Merris conjecture, ex-
cept that Lemma 3.8.3 still has to be proved.
Proof of Lemma 3.8.3

Let Γ be a semi-bipartite graph with clique of size c and coclique of size n − c.
The
partition of
the vertex set induces a partition of the Laplace matrix L =
K + D −A
, where K is the Laplacian of the complete graph Kc and A is
−A⊤ E
the c × (n − c) adjacency matrix between vertices in the clique and the coclique,
and D and E are diagonal matrices with the row and column sums of A.
Pc
Step 5 If νc ≥ c, then i=1 #{x | dx ≥ i} = c2 + tr D.
Indeed, all vertices in the clique have degree at least c, for if some vertex x
in the clique had degree c − 1 then we could move it to the cocliquePcand find
νc ≤ c − 1 P
from Step 3, contrary to the assumption. It follows that i=1 #{x |
dx ≥ i} = x min(c, dx ) = c2 + tr E = c2 + tr D.
Step 6 Suppose that the subspace W spanned by
the L-eigenvectors
belonging

I I I
to ν1 , . . . , νc is spanned by the columns of . Then L = Z for
X X X
Pc
some matrix Z, and i=1 νi = tr Z.

U U U
Indeed, if has these eigenvectors as columns, then L = T where
V V V

U I
T is the diagonal matrix with the eigenvalues. Now = U , so that
V X
I I Pc
L = Z where Z = U T U −1 and tr Z = tr T = i=1 νi .
X X
Suppose we are in the situation of the previous step, and that moreover X is
nonpositive. Let δ be the maximum degree among the vertices in the coclique,
2
so that δ ≤c. We to show that if νc > c or νc = c > δ then tr Z ≤ c +tr D.
have
I K + D − AX
Now L = , so Z = K + D − AX, and tr Z = tr(K +
X −A⊤ + EX
D − AX) = c(c − 1) + tr D − tr(AX), and we need tr(AX) ≥ −c. But since
c < n, the eigenvectors are orthogonal to 1 so that X has column sums −1.
Since X is nonpositive, tr(AX) ≥ −c follows, and we are done.
By interlacing νc+1 is at most the largest eigenvalue of E, that
is δ, which
0
by hypothesis is smaller than νc . Hence the subpace of vectors meets W
∗
trivially, so that W has a basis of the required form. Only nonpositivity of X
remains, and the following lemma completes the proof.
Lemma 3.8.4 If νc > δ, then the invariant subspace W spanned by the L-

I
eigenvectors for νi , 1 ≤ i ≤ c, is spanned by the columns of where X is
X
3.9. THE LAPLACIAN FOR HYPERGRAPHS 53
nonpositive.
Proof. We argue by continuity, viewing L = L(A) and X = X(A) as functions

of the real-valued matrix A, where 0 ≤ A ≤ J. (Now D has the row sums of A,
and E has the column sums, and δ is the largest element of the diagonal matrix
E.) We write J for the c × (n − c) all-1 matrix, and Jc for the all-1 matrix of
order c, so that JX = −Jc .
Our hypothesis νc > δ holds for all matrices L(α) := L(αA + (1 − α)J) =
(α)
αL + (1 − α)L(0) , for 0 ≤ α ≤ 1. Indeed, let L(α) have eigenvalues νi , so that
(0) (0) (0)
νc = n and νc+1 = νn−1 = c. The matrix L(α) has lower left-hand corner
αE + (1 − α)cI so that δ (α) = αδ + (1 − α)c. The c-space W is orthogonal to
(α) (α)
1, so that νc ≥ ανc + (1 − α)c (by Theorem 2.4.1), and hence νc > δ (α) for
(0)
0 < α ≤ 1, and also for α = 0, since νc = n and δ (0) = c. It follows that
(α) (α)
νc > νc+1 for 0 ≤ α ≤ 1.
As we used already, L(J) has spectrum nc , cn−c−1 , 0, and one checks that
1
X(J) = − n−c J ⊤ < 0, as desired. Above we found the condition XZ = −A⊤ +
EX on X, that is, X(K + D − AX) − EX + A⊤ = 0, that is, X(K + Jc + D) =
XJc +XAX +EX −A⊤ = −X(J −A)X +EX −A⊤ . It follows, since K +Jc +D
is a positive diagonal matrix, that if X ≤ 0 and A > 0, then X < 0. The matrix
X(A) depends continuously on A (in the region where νc+1 < νc ) and is strictly
negative when A > 0. Then it is nonpositive when A ≥ 0.
3.9 The Laplacian for hypergraphs

Let a simplicial complex on a finite set S be a collection C of subsets of S (called
simplices) that is an order ideal for inclusion, that is, such that if A ∈ C and
B ⊆ A then also B ∈ C. Let the dimension of a simplex A be one less than
its cardinality, and let the dimension of a simplicial complex be the maximum
of the dimensions of its simplices. Given a simplicial complex C, let Ci (for
i ≥ −1) be the vector space (over any field) that has the simplices of dimension
i as basis. Order the simplices arbitrarily
P (say, using some order on S) and
define ∂i : Ci → Ci−1 by ∂i s0 . . . si = j (−1)j s0 . . . sbj . . . si . Then ∂i−1 ∂i = 0
for all i ≥ 0.
⊤
Let Ni be the matrix of ∂i on the standard basis, and put Li = Ni+1 Ni+1
and L′i = Ni⊤ Ni . The matrices Li generalize the Laplacian. Indeed, in the case
of a 1-dimensional simplicial complex (that is, a graph) the ordinary Laplace
matrix is just L0 , and L′0 is the all-1 matrix J.
Since ∂i ∂i+1 = 0 we have Li L′i = L′i Li = 0, generalizing LJ = JL = 0.
We have tr Li−1 = tr L′i = (i + 1)|Ci |. This generalizes the facts that tr L is
twice the number of edges, and tr J the number of vertices.
In case the underlying field is R, we have the direct sum decomposition
Ci = im Ni+1 ⊕ ker(Li + L′i ) ⊕ im Ni⊤ . (Because then M ⊤ M x = 0 if and only
if M x = 0.) Now ker Ni = im Ni+1 ⊕ ker(Li + L′i ) so that the i-th reduced
homology group is Hi (C) := ker Ni /im Ni+1 ∼ = ker(Li + L′i ).
Example The spectrum of Lm−2 for a simplicial complex containing all m-
subsets of an n-set (the complete
m-uniform hypergraph) consists of the eigen-
n−1
value n with multiplicity m−1 and all further eigenvalues are 0.
Indeed, we may regard simplices s0 . . . sm−1 as elements s0 ∧ . . . ∧ sm−1 of an

exterior algebra. Then the expression s0 . . . sm−1 is defined regardless of the
order
P of the factors, and also when factors are repeated. Now Ni t0 . . . ti =
j (−1) j
t0 . . . tbj . . . ti and for the complete (i + 2)-uniform hypergraph we have
P
Ni t0 . . . ti = t tt0 . . . ti , so that Li +L′i = nI. It follows that Ni+1
⊤ ⊤ ⊤
Ni+1 Ni+1 =
⊤
nNi+1 , and Li has eigenvalues 0 and n. The multiplicities follow by taking the
trace.
Duval & Reiner [136] generalized the Grone-Merris conjecture. Given an m-
uniform hypergraph H, let dx be the number of edges containing the vertex x.
Let the spectrum of H be that of the matrix Lm−2 for the simplicial complex
consisting of all subsets of edges of H.
Conjecture Let the m-uniform hypergraph H have degrees dx , and Laplace
eigenvalues νi , ordered such that ν1 ≥ ν2 ≥ . . . ≥ 0. Then for all t we have
t
X t
X
νj ≤ #{x | dx ≥ j}.
j=1 j=1
Equality for all t holds if and only if H is invariant under downshifting.

The part about ‘downshifting’ means the following: Put a total order on the
vertices of H in such a way that if x ≤ y then dx ≥ dy . Now H is said to
be invariant under downshifting if whenever {x1 , . . . , xm } is an edge of H, and
{y1 , . . . , ym } is an m-set with yi ≤ xi for all i, then also {y1 , . . . , ym } is an edge
of H. If this holds for one total order, then it holds for any total order that is
compatible with the degrees.
For m = 2 this is precisely the Grone-Merris conjecture. (And the graphs that
are invariant for downshifting are precisely the threshold graphs.) The ‘if’ part
of the equality case is a theorem:
Theorem 3.9.1 (Duval & Reiner [136]) If H is an m-uniform hypergraph with

degrees dx and Laplace eigenvalues νi with ν1 ≥ ν2 ≥ . . . ≥ 0, and H is invariant
for downshifting, then νj = #{x | dx ≥ j} for all t.
In particular it follows that hypergraphs invariant for downshifting have

integral Laplace spectrum.
For example, the complete
m-uniform hypergraph on an underlying set of
n−1 n−1
size n has degrees m−1 so that νj = n for 1 ≤ j ≤ m−1 and νj = 0 for
n−1
n

m−1 < j ≤ m , as we already found earlier.
Dominance order
The conjecture and theorem can be formulated more elegantly in terms of dom-
inance order. Let a = (ai ) and b = (bi ) be two finite nonincreasing sequences of
nonnegative
Pt Preal numbers. We say P∞that b dominates
P∞ a, and write a E b, when
t
i=1 ai ≤ i=1 bi for all t, and i=1 ai = i=1 bi , where missing elements are
taken to be zero.
For example, in this notation Schur’s inequality (Theorem 2.6.1) says that
d E θ if d is the sequence of diagonal elements and θ the sequence of eigenvalues
of a real symmetric matrix.
3.10. APPLICATIONS OF EIGENVECTORS 55
If a = (aj ) is a finite nonincreasing sequences of nonnegative integers, then

a⊤ denotes the sequence (a⊤ ⊤
j ) with aj = #{i | ai ≥ j}. If a is represented by a
Ferrers diagram, then a⊤ is represented by the transposed diagram.
For example, the Duval-Reiner conjecture says that µ′ E d⊤ .
If a and b are two nonincreasing sequences, then let a ∪ b denote the (mul-
tiset) union of both sequences, with elements sorted in nonincreasing order.
Lemma 3.9.2
(i) a⊤⊤ = a,
(ii) (a ∪ b)⊤ = a⊤ + b⊤ and (a + b)⊤ = a⊤ ∪ b⊤ ,
(iii) a E b if and only if b⊤ E a⊤ .
3.10 Applications of eigenvectors

Sometimes it is not the eigenvalue but the eigenvector that is needed. We sketch
very briefly some of the applications.
3.10.1 Ranking
0.15 0.08
• •
0.42 0.52 0.27 0.08

• • • • θ = 3.35
• • •
0.42 0.47 0.22
Figure 3.1: Graph with Perron-Frobenius eigenvector
In a network, important people have many connections. One would like to

pick out the vertices of highest degree and call them the most important. But
it is not just the number of neighbors. Important people have connections to
many other important people. If one models this and says that up to some
constant of proportionality one’s importance is the sum of the importances of
one’s neighbors in the graph, then the vector giving the importance of each
vertex becomes an eigenvector of the graph, necessarily the Perron-Frobenius
eigenvector if importance cannot be negative. The constant of proportionality
is then the largest eigenvalue.
3.10.2 Google Page rank

Google uses a similar scheme to compute the Page Rank [40] of web pages. The
authors described (in 1998) the algorithm as follows:
Suppose pages x1 , ..., xm are the pages that link to a page y. Let
page xi have di outgoing links. Then the PageRank of y is given by
X P R(xi )
P R(y) = 1 − α + α .
i
di
P
The PageRanks form a probability distribution: x P R(x) = 1. The
vector of PageRanks can be calculated using a simple iterative algo-
rithm, and corresponds to the principal eigenvector of the normalized
link matrix of the web. A PageRank for 26 million web pages can be
computed in a few hours on a medium size workstation. A suitable
value for α is α = 0.85.
In other words, let Γ be the directed graph on n vertices consisting of all

web pages found, with an arrow from x to y when page x contains a hyperlink
to page y. Let A be the adjacency matrix of Γ (with Axy = 1 if there is a link
from x to y). Let D be the diagonal matrix of outdegrees, so that the scaled
matrix S = D−1 A has row sums 1, and construct the positive linear combination
M = 1−α n J + αS with 0 < α < 1. Since M P > 0 the matrix M has a unique
positive left eigenvector u, normed so that ux = 1. Now M 1 = 1 and hence
uM = u. The PageRank of the web page x is the value ux .
A small detail is the question what to do when page x does not have outgoing
edges, so that row x in A is zero. One possibility is to do nothing (and take
Dxx = 1). Then u will have eigenvalue less than 1.
The vector u is found by starting with an approximation (or just any positive
vector) u0 and then computing the limit of the sequence ui = u0 M i . That is
easy: the matrix M is enormous, but A is sparse: on average a web page
does not have more than a dozen links. The constant α regulates the speed of
convergence: convergence is determined by the 2nd largest eigenvalue, which
is bounded by α ([211]). It is reported that 50 to 100 iterations suffice. A
nonzero α guarantees that the matrix is irreducible. An α much less than 1
guarantees quick convergence. But an α close to 1 is better at preserving the
information in A. Intuitively, ux represents the expectation of finding oneself at
page x after many steps, where each step consists of either (with probability α)
clicking on a random link on the current page, or (with probability 1−α) picking
a random internet page. Note that the precise value of ux is unimportant—only
the ordering among the values ux is used.
There are many papers (and even books) discussing Google’s PageRank. See
e.g. [66], [24].
3.10.3 Cutting
Often the cheapest way to cut a connected graph into two pieces is by partition-
ing it into a single vertex (of minimal valency) and the rest. But in the area of
clustering (see also below) one typically wants relatively large pieces. Here the
second Laplace eigenvector helps. Without going into any detail, let us try the
same example as above.
We see that cutting the edges where the 2nd Laplace eigenvector changes
sign is fairly successful in this case.
3.10.4 Graph drawing

Often, a reasonable way to draw a connected graph is to take Laplace eigen-
vectors u and v for the 2nd and 3rd smallest Laplace eigenvalues, and draw the
vertex x at the point with coordinates (ux , vx ). See, e.g. [218].
3.10. APPLICATIONS OF EIGENVECTORS 57
−0.41 0.52
• •
−0.27 −0.20 0.25 0.52

• • • • µ2 = 0.51
• • •
−0.27 −0.20 0.04
Figure 3.2: Graph with 2nd Laplace eigenvector
One can justify this as follows. Let thePenergy of an embedding ρ : Γ → Rm

be the sum of the squared edge lengths e ||ρ(x) − ρ(y)||2 where the sum is
over all edges e = xy. Let R be the m × n matrix of which the columns are
the vertex images ρ(x). Then the energy of ρ equals RLR⊤ . For graph drawing
one would like to minimize the energy, given some normalization so that not
all vertices are mapped close to the origin or close to some lower-dimensional
subspace of Rm . Pisanski & Shawe-Taylor [260] propose to require R1 = 0 and
RR⊤ = I, so that the origin is the center of mass, and ||R⊤ v||2 = ||v||2 for all
vectors v ∈ Rm : no vector is almost P perpendicular to the entire drawing. In
m+1
this situation the minimum energy is i=2 µi , and this minimum is achieved
when the row space of R contains the Laplace eigenvectors of µ2 , . . . , µm+1 . The
authors also discuss variations of this setup.
3.10.5 Clustering
Given a large data set, one often wants to cluster it. If the data is given as a
set of vectors in some Euclidean space Rm , then a popular clustering algorithm
is k-means:
Given a set X of N vectors inPRm and P a number k, 2find a partition of X into
k
k subsets X1 , . . . , XP
k such that i=1 x∈Xi ||x − ci || is as small as possible,
where ci = (1/|Xi |) x∈Xi x is the centroid of Xi .
The usual algorithm uses an iterative approach. First choose the k vectors
ci in some way, arbitrary or not. Then take Xi to be the subset of X consisting
of the vectors closer to ci than to the other cj (breaking ties arbitrarily). Then
compute new vectors ci as the centroids of the sets Xi , and repeat. In common
practical situations this algorithm converges quickly, but one can construct ex-
amples where this takes exponential time. The final partition found need not
be optimal, but since the algorithm is fast, it can be repeated a number of times
with different starting points ci .
Now if the data is given as a graph, one can compute eigenvectors u1 , . . . , um
for the m smallest eigenvalues µ1 , . . . , µm of the Laplace matrix L, and assign
to the vertex x the vector (ui (x))i and apply a vector space clustering algorithm
such as k-means to the resulting vectors.
This is reasonable. For example, if the graph is disconnected with c con-
nected components, then the first c eigenvalues of L are zero, and the first
c eigenvectors are (linear combinations of) the characteristic functions of the
connected components.
This approach also works when one has more detailed information—not ad-
jacent/nonadjacent but a (nonnegative) similarity or closeness P measure. (One
uses an edge-weighted graph, with Axy = w(x, y) and dx = y w(x, y) and
D the diagonal matrix with P Dxx = dx , and L = D − A. Again L is positive
semidefinite, with u⊤ Lu = w(x, y)(u(x) − u(y))2 . The multiplicity of the
eigenvalue 0 is the number of connected components of the underlying graph
where points x, y are adjacent when w(x, y) > 0.)
Especially important is the special case where one searches for the cheapest
cut of the graph into two relatively large pieces. If the graph is connected, then
map the vertices into R1 using x 7→ u(x), where u is the eigenvector for the
second smallest eigenvalue of L, and then use 2-means to cluster the resulting
points. Compare §1.7 on the algebraic connectivity of a graph.
Several matrices related to the Laplacian have been used in this context. It
seems useful to normalize the matrix, and to retain the property that if the graph
is disconnected the characteristic functions of components are eigenvectors. A
suitable matrix is Lnorm = D−1 L = I − D−1 A.
There is a large body of literature on clustering in general and spectral
clustering in particular. A few references are [169, 235, 288].
3.10.6 Searching an eigenspace

There exists a unique strongly regular graph1 with parameters (v, k, λ, µ) =
(162, 56, 10, 24) found as the second subconstituent of the McLaughlin graph.
Its vertex set can be split into two halves such that each half induces a strongly
regular graph with parameters (v, k, λ, µ) = (81, 20, 1, 6). How many such splits
are there? Can we find them all?
In this and many similar situations one can search an eigenspace. The first
graph has spectrum 561 2140 (−16)21 and a split gives an eigenvector with
eigenvalue −16 if we take the vector that is 1 on the subgraph and −1 on the
rest.
It is easy to construct an explicit basis (ui ) for the 21-dimensional eigenspace,
where the j-th coordinate of ui is δij . Construct the 221 eigenvectors that are
±1 on the first 21 coordinates and inspect the remaining coordinates. If all are
±1 one has found a split into two regular graphs of valency 20. In this particular
case there are 224 such subgraphs, 112 splits, and all subgraphs occurring are
strongly regular with the abovementioned parameters.
3.11 Exercises
Exercise 1 Consider a graph with largest√ eigenvalue θ1 and maximum valency
kmax . Use interlacing to show that θ1 ≥ kmax . When does equality hold?
Exercise 2 Let Γ be a k-regular graph with n vertices and eigenvalues k =
θ1 ≥ . . . ≥ θn . Let Γ′ be an induced subgraph of Γ with n′ vertices and average
degree k ′ .
nk′ −n′ k
(i) Prove that θ2 ≥ n−n′ ≥ θn .
1 For strongly regular graphs, see Chapter 8. No properties are used except that the sub-
structure of interest corresponds to an eigenvector of recognizable shape.

3.11. EXERCISES 59
(ii) What can be said in case of equality (on either side)?
(iii) Deduce Hoffman’s bound (Theorem 4.1.2) from the above inequality.
Exercise 3 Prove the conjecture from Section 3.8.1 for regular graphs. (Hint:
use Cauchy-Schwarz.)
Chapter 4
Substructures
Interlacing yields information on subgraphs of a graph, and the way such sub-
graphs are embedded. In particular, one gets bounds on extremal substructures.
Probably the most important use of the spectrum is its use to imply con-
nectivity, randomness, and expansion properties of a graph.
For the adjacency spectrum the important parameter is the second largest
eigenvalue θ2 , cf. §1.7, or rather the ratio θ2 /θ1 . A small ratio implies high
connectivity. For the Laplace spectrum it is the second smallest, µ2 , or rather
the ratio µ2 /µn . A large ratio implies high connectivity.
The Laplace eigenvector belonging to µ2 has been used in graph partitioning
problems.
4.1 Cliques and cocliques

A clique in a graph is a set of pairwise adjacent vertices. A coclique in a graph
is a set of pairwise nonadjacent vertices. The clique number ω(Γ) is the size of
the largest clique in Γ. The independence number α(Γ) is the size of the largest
coclique in Γ.
Let Γ be a graph on n vertices (undirected, simple, and loopless) having an
adjacency matrix A with eigenvalues θ1 ≥ . . . ≥ θn . Both Corollaries 2.5.2 and
2.5.4 lead to a bound for α(Γ).
Theorem 4.1.1 α(Γ) ≤ n−n− = |{i|θi ≥ 0}| and α(Γ) ≤ n−n+ = |{i|θi ≤ 0}|.
Proof. A has a principal submatrix B = 0 of size α = α(Γ). Corollary 2.5.2
gives θα ≥ ηα = 0 and θn−α−1 ≤ η1 = 0.
Theorem 4.1.2 If Γ is regular of nonzero degree k, then
−θn
α(Γ) ≤ n ,
k − θn
and if a coclique C meets this bound, then every vertex not in C is adjacent to
precisely −θn vertices of C.
Proof. We apply Corollary 2.5.4. The coclique gives rise to a partition of A
with quotient matrix
0 k
B= kα kα ,
n−α k − n−α
61
62 CHAPTER 4. SUBSTRUCTURES
where α = α(Γ). B has eigenvalues η1 = k = θ1 (the row sum) and η2 =

−kα/(n − α) (since trace B = k + η2 ) and so θn ≤ η2 gives the required
inequality. If equality holds, then η2 = θn , and since η1 = θ1 , the interlacing is
tight and hence the partition is equitable.
The first bound is due to Cvetković [103]. The second bound is an unpublished
result of Hoffman.
Example The Petersen graph has 10 vertices and spectrum 31 15 (−2)4 , and
its independence number is 4. So both bounds are tight.
4.1.1 Using weighted adjacency matrices

Let us call a real symmetric matrix B a weighted adjacency matrix of a graph Γ
when B has rows and columns indexed by the vertex set of Γ, has zero diagonal,
and satisfies Bxy = 0 whenever x 6∼ y.
The proof of Theorem 4.1.1 applies to B instead of A, and we get
Theorem 4.1.3 α(Γ) ≤ n − n− (B) and α(Γ) ≤ n − n+ (B).
Similarly, the proof of Theorem 4.1.2 remains valid for weighted adjacency
matrices B with constant row sums.
Theorem 4.1.4 Let B be a weighted adjacency matrix of Γ with constant row

sums b and smallest eigenvalue s. Then α(Γ) ≤ n(−s)/(b − s).
4.2 Chromatic number

A proper vertex coloring of a graph is an assignment of colors to the vertices so
that adjacent vertices get different colors. (In other words, a partition of the
vertex set into cocliques.) The chromatic number χ(Γ) is the minimum number
of colors of a proper vertex coloring of Γ.
Proposition 4.2.1 (Wilf [307]) Let Γ be connected with largest eigenvalue θ1 .

Then χ(Γ) ≤ 1 + θ1 with equality if and only if Γ is complete or is an odd cycle.
Proof. Put m = χ(Γ). Since Γ cannot be colored with m − 1 colors, while

coloring vertices of degree less than m − 1 is easy, there must be an induced
subgraph ∆ of Γ with minimum degree at least m − 1. Now θ1 ≥ θ1 (∆) ≥
dmin (∆) ≥ m − 1 = χ(Γ) − 1. If equality holds, then by Perron-Frobenius Γ = ∆
and ∆ is regular of degree m − 1 (by Proposition 3.1.2), and the conclusion
follows by Brooks’ theorem.
Since each coclique (color class) has size at most α(Γ), we have χ(Γ) ≥ n/α(Γ)
for a graph Γ with n vertices. Thus upper bounds for α(Γ) give lower bounds for
χ(Γ). For instance if Γ is regular of degree k = θ1 then Theorem 4.1.2 implies
that χ(Γ) ≥ 1 − θθn1 . This bound remains however valid for non-regular graphs.
θ1
Theorem 4.2.2 (Hoffman [195]) If Γ is not edgeless then χ(Γ) ≥ 1 − .
θn
Proof. Put m = χ(Γ). Since Γ is not edgeless, θn < 0. Now, by part (i) of
the following proposition, θ1 + (m − 1)θn ≤ θ1 + θn−m+2 + . . . + θn ≤ 0.
4.2. CHROMATIC NUMBER 63
Proposition 4.2.3 Put m = χ(Γ). Then
(i) θ1 + θn−m+2 + . . . + θn ≤ 0.
(ii) If n > m, then θ2 + . . . + θm + θn−m+1 ≥ 0.
(iii) If n > tm, then θt+1 + . . . + θt+m−1 + θn−t(m−1) ≥ 0.
Proof. Let A have orthonormal eigenvectors uj , so that Auj = θj uj .

(i) Let {X1 , . . . , Xm } be a partition of Γ into m cocliques, where m = χ(Γ).
Let xPj be the pointwise product of u1 with the characteristic vector of Xj , so
that xj = u1 . Now apply Corollary 2.5.3 to the vectors xj , after deleting
those that are zero. The matrix C defined there satisfies C1 = θ1 1, and has
zero diagonal, and has eigenvalues ηj interlacing those of A. Hence
0 = tr(C) = η1 + . . . + ηm ≥ θ1 + θn−m+2 + . . . + θn .
(ii) Put A′ = A − (θ1 − θn )u1 u⊤ ′

1 , then A has the same eigenvectors uj as A,
but with eigenvalues θn , θ2 , ..., θn . Pick a non-zero vector y in
⊥
hun−m+1 , . . . , un i ∩ hx1 , . . . , xm i .
The two spaces have non-trivial intersection since the dimensions add up to n
and u1 is orthogonal to both. Let yjPbe the pointwise product of y with the
characteristic vector of Xj , so that yj = y and yj⊤ A′ yj = 0. Now apply
′
Corollary 2.5.3 to the matrix A and the vectors yj , after deleting those that
are zero. The matrix C defined there has zero diagonal, and smallest eigenvalue
⊤
smaller than the Rayleigh quotient yy⊤Ay
y
, which by choice of y is at most θn−m+1 .
We find
0 = tr(C) = η1 + . . . + ηm ≤ θ2 + θ3 + . . . + θm + θn−m+1 .
(iii) The proof is as under (ii), but this time we move t (instead of just
one) eigenvalues away (by subtracting multiples of uj u⊤ j for 1 ≤ j ≤ t). The
vector y must be chosen orthogonal to tm vectors, which can be done inside
the (tm − t + 1)-space hun−tm+t , . . . , un i, assuming that this space is already
orthogonal to u1 , ..., ut , i.e., assuming that n > tm.
The above proof of Theorem 4.2.2 using (i) above appeared in [174].
A coloring that meets the bound of Theorem 4.2.2 is called a Hoffman col-
oring. For regular graphs, the color classes of a Hoffman coloring are cocliques
that meet Hoffman’s coclique bound. So in this case all the color classes have
equal size and the corresponding matrix partition is equitable.
In [175] more inequalities of the above kind are given. But the ones men-
tioned here, especially (i) and (ii), are by far the most useful.
Example The complete multipartite graph Km×a has chromatic number m and
spectrum (am − a)1 0m(a−1) (−a)m−1 . It has equality in Hoffman’s inequality
(and hence in (i)), and also in (ii).
Example The graph obtained by removing √ an edge from Kn has √ chromatic
number n − 1 and spectrum 12 (n − 3 + D), 0, (−1)n−3 , 21 (n − 3 − D) with
D = (n + 1)2 − 8, with equality in (i).
Example Consider the generalized octagon of order (2, 4) on 1755 vertices. It

has spectrum 101 5351 1650 (−3)675 (−5)78 . It is not 3-chromatic, as one sees
by removing the largest 352 eigenvalues, i.e., by applying (iii) with t = 352.
The inequality (ii) looks a bit awkward, but can be made more explicit if
the smallest eigenvalue θn has a large multiplicity.
Corollary 4.2.4 If the eigenvalue θn has multiplicity m and θ2 > 0, then
θn
χ(Γ) ≥ min(1 + m, 1 − ).
θ2
θn
Proof. If χ(Γ) ≤ m, then θn = θn−χ(Γ)+1 , hence χ(Γ) ≥ 1 − θ2 .
4.2.1 Using weighted adjacency matrices

If Γ has an m-coloring, then Γ × Km has an independent set of size n, the
number of vertices of Γ. This means that one can use bounds on the size of an
independent set to obtain bounds on the chromatic number.
Example Consider the generalized octagon of order (2, 4) again. Call it Γ,
and call its adjacency matrix A. Now consider the weighted adjacency matrix
B of K3 × Γ, where the K3 is weighted with some number r, where 1 < r < 23 .
For each eigenvalue θ of A, we find eigenvalues θ + 2r (once) and θ − r (twice)
as eigenvalues of B. Applying Theorem 4.1.3 we see that α(K3 × Γ) ≤ 3(1 +
351) + 650 = 1706 while Γ has 1755 vertices, so Γ is not 3-chromatic.
4.2.2 Rank and chromatic number

The easiest way for A to have low rank, is when it has many repeated rows. But
then Γ contains large cocliques. People have conjectured that it might be true
that χ(Γ) ≤ rk A when A 6= 0. A counterexample was given by Alon & Seymour
[7] who observed that the complement of the folded 7-cube (on 64 vertices) has
chromatic number χ = 32 (indeed, α = 2), and rank 29 (indeed, the spectrum
of the folded 7-cube is 71 321 (−1)35 (−5)7 ).
4.3 Shannon capacity

Shannon [282] defined the Shannon capacity of a graph as the capacity of the
zero-error channel where a transmission consists of sending a vertex of the graph,
and two transmissions can be confused when the corresponding vertices are
joined by an edge.
The maximum number of possible messages of length 1 is α(Γ), and log α(Γ)
bits are transmitted. The maximum number of possible messages of length
ℓ is α(Γℓ ) where Γℓ is defined (here) as the graph on sequences of ℓ ver-
tices from Γ, where two sequences are adjacent when on each coordinate po-
sition their elements are equal or adjacent. This explains the definition c(Γ) =
supℓ→∞ α(Γℓ )1/ℓ . Now the channel capacity is log√c(Γ).
For example, for the pentagon we find c(Γ) ≥ 5 as shown by the 5-coclique
00, 12, 24, 31, 43.
Computing c(Γ) is a difficult unsolved problem, even for graphs as simple as
C7 , the 7-cycle.
4.4. STARS AND STAR COMPLEMENTS 65
Clearly, α(Γ) ≤ c(Γ) ≤ χ(Γ). (Indeed, if m = χ(Γ) then Γ can be covered

by m cliques, and Γℓ can be covered by mℓ cliques, and α(Γℓ ) ≤ mℓ .) In a few
cases this suffices to determine c(Γ).
One can sharpen the upper bound to the fractional clique covering number.
For example, the vertices of C5 can be doubly covered by 5 cliques, so the
vertices of C5ℓ can be covered 2ℓ times by 5ℓ cliques, and α(C5ℓ ) ≤ (5/2)ℓ so that
c(C5 ) ≤ 5/2.
If A is the adjacency matrix of Γ, then ⊗ℓ (A + I) − I is the adjacency matrix
of Γℓ .
The Hoffman upper bound for the size of cocliques is also an upper bound
for c(Γ) (and therefore, when the Hoffman bound holds with equality, also the
Shannon capacity is determined).
Proposition 4.3.1 (Lovász [232]) Let Γ be regular of valency k. Then
c(Γ) ≤ n(−θn )/(k − θn ).
Proof. Use the weighted Hoffman bound (Theorem 4.1.4). If B = A − θn I,

then ⊗ℓ B − (−θn )ℓ I has constant row sums (k − θn )ℓ − (−θn )ℓ and smallest
eigenvalue −(−θn )ℓ , so that α(Γℓ ) ≤ (n(−θn )/(k − θn ))ℓ .
√ √
Using n = 5, k = 2, θn = (−1 − 5)/2 we find for the pentagon c(Γ) ≤ 5.
Hence equality holds.
Haemers [172, 173] observed that if B is a matrix indexed by the vertices of Γ
and Bxx 6= 0 for all x, and Bxy = 0 whenever x 6∼ y, then c(Γ) ≤ rk B. Indeed,
for such a matrix α(Γ) ≤ rk B since an independent set determines a submatrix
that is zero outside a nonzero diagonal. Now rk ⊗ℓ B = (rk B)ℓ .
Note that in the above the rank may be taken over any field.
Example The collinearity graph Γ of the generalized quadrangle GQ(2, 4) (the
complement of the Schläfli graph, cf. §8.6) on 27 vertices has spectrum 101 120
(−5)6 . Taking B = A − I shows that c(Γ) ≤ 7. (And c(Γ) ≥ α(Γ) = 6.) The
complement Γ has α(Γ) = 3, but this is √ also the Hoffman bound, so c(Γ) = 3.
Alon [3] proves that c(Γ + Γ) ≥ 2 n for all Γ. Combined with the above
example, this shows that the Shannon capacity of the disjoint sum of two graphs
can be larger than the sum of their Shannon capacities.
4.4 Stars and star complements

Consider a graph Γ with vertex set X. By interlacing, the multiplicity of any
given eigenvalue changes by at most 1 if we remove a vertex. But there is
always a vertex such that removing it actually decreases the multiplicity. And
that means that if θ is an eigenvalue of multiplicity m we can find a star subset
for θ, that is, a subset S of X of size m such that Γ \ S does not have eigenvalue
θ. Now X \ S is called a star complement.
Why precisely can we decrease the multiplicity? Let u be a θ-eigenvector of
A, so that (θI − A)u = 0, and let x be a vertex with ux 6= 0. Then removing x
from Γ decreases the multiplicity of θ.
Indeed, removing x is equivalent to the two P actions: (i) forcing ux = 0 for
eigenvectors u, and (ii) omitting the condition y ∼ x uy = θux (row x of the
matrix equation (θI − A)u = 0) for eigenvectors u. Since A is symmetric, the

column dependency (θI − A)u = 0 given by u is also a row dependency, and row
x is dependent on the remaining rows, so that (ii) doesn’t make a difference.
But (i) does, as the vector u shows. So the multiplicity goes down.
This argument shows that the star sets for θ are precisely the sets S of size
m such that no θ-eigenvector vanishes on all of S. Also, that any subgraph
without eigenvalue θ is contained in a star complement.
Proposition 4.4.1 ([137, 108]) Let Γ be a graph with eigenvalue θ of multi-

plicity m. Let S be a subset of the vertex
setX of Γ, and let the partition
B C
{S, X \ S} of X induce a partition A = of the adjacency matrix A. If
C⊤ D
S is a star set for θ (i.e., if |S| = m and D does not have eigenvalue θ), then
B − θI = C(D − θI)−1 C ⊤ .
Proof. The row space of A − θI has rank n − m. If S is a star set, then

this row space is spanned by the rows of [C ⊤ D − θI]. Alternatively, apply
Corollary 2.7.2 to A − θI.
This proposition says that the edges inside a star set are determined by the rest
of the graph (and the value θ). Especially when m is large, this may be useful.
Stars and star complements have been used to study exceptional graphs
with smallest eigenvalue not less than −2, see, e.g., [107, 109, 110]. (One starts
with the observation that if θ is the smallest eigenvalue of a graph, then a star
complement has smallest eigenvalue larger than θ. But all graphs with smallest
eigenvalue larger than −2 are explicitly known.) Several graphs and classes of
graphs have been characterized by graph complement. See, e.g., [207, 109].
A star partition is a partition of X into star sets Sθ for θ, where θ runs
through the eigenvalues of Γ. It was shown in [108] that every graph has a star
partition.
4.5 The second largest eigenvalue

There is a tremendous amount of literature about the second largest eigenvalue
of a regular graph. If the gap between the largest and second largest eigenvalues
is large, then the graph has good connectivity, expansion and randomness prop-
erties. (About connectivity, see also §1.7.)
4.5.1 Bounds for the second largest eigenvalue

In this connection it is of interest how large this gap can become. Theorems by
Alon-Boppana
√ and Serre say that for large graphs θ2 cannot be much smaller
than 2 k − 1.
Proposition 4.5.1 (Alon-Boppana [2]) If k ≥ 3 then for k-regular graphs on

n vertices one has
√ log(k − 1)
θ2 ≥ 2 k − 1 (1 − O( )).
log n
4.5. THE SECOND LARGEST EIGENVALUE 67
Proposition 4.5.2 (Serre [281]) For each ǫ > 0, there exists a positive constant
c = c(ǫ, k) such that for any k-regular
√ graph Γ on n vertices, the number of
eigenvalues of Γ larger than (2 − ǫ) k − 1 is at least cn.
Quenell gives (weaker) explicit bounds:
Proposition 4.5.3 ([262]) Let Γ be a finite graph with diameter d and minimal
degree k ≥ 3. Then for 2 ≤ m √≤ 1 + d/4, the mth eigenvalue of the adjacency
π
matrix A of Γ satisfies λm > 2 k − 1 cos( r+1 ), where r = ⌊d/(2m − 2)⌋.
Alon [2] conjectured, and Friedman [147] proved √ that large random k-regular
graphs have second largest eigenvalue smaller than 2 k − 1+ǫ (for fixed k, ǫ > 0
and n sufficiently large). Friedman remarks that numerical experiments
√ seem
to indicate that random k-regular graphs in fact satisfy θ2 < 2 k − 1. √
A connected k-regular graph is called a Ramanujan graph when |θ| ≤ 2 k − 1
for all eigenvalues θ 6= k. (This notion was introduced in [234].) It is not difficult
to find such graphs. For example, complete graphs, or Paley graphs, will do.
Highly nontrivial was the construction of infinite sequences of Ramanujan graphs
with given, constant, valency k and size n tending to infinity. Lubotzky, Phillips
& Sarnak [234] and Margulis [240] constructed for each prime p ≡ 1 (mod 4) an
infinite series of Ramanujan graphs with valency k = p + 1.
4.5.2 Large regular subgraphs are connected

We note the following trivial but useful result,
Proposition 4.5.4 Let Γ be a graph with second largest eigenvalue θ2 . Let ∆

be a nonempty regular induced subgraph with largest eigenvalue ρ > θ2 . Then ∆
is connected.
Proof. The multiplicity of the eigenvalue ρ of ∆ is the number of connected

components of ∆, and by interlacing this is 1.
4.5.3 Randomness
Let Γ be a regular graph of valency k on n vertices, and assume that (for some
real constant λ) we have |θ| ≤ λ for all eigenvalues θ 6= k. The ratio λ/k
determines randomness and expansion properties of Γ: the smaller λ/k, the
more random, and the better expander Γ is.
For example, the following proposition says that most points have approx-
imately the expected number of neighbors in a given subset of the vertex set.
Here Γ(x) denotes the set of neighbours of the vertex x in the graph Γ.
Proposition 4.5.5 Let R be a subset of size r of the vertex set X of Γ. Then

X kr 2 r(n − r) 2
(|Γ(x) ∩ R| − ) ≤ λ .
n n
x∈X
Proof. Apply interlacing to A2 and the partition {R, X \ R} of X. The sum

of all entries of the matrix A2 in the (R, R)-block
P equals the number of paths
y ∼ x ∼ z, with y, z ∈ R and x ∈ X, that is, x (|Γ(x) ∩ R|)2 .
Rather similarly, the following proposition, a version of the expander mixing

lemma from Alon & Chung [5], says that there are about the expected number
of edges between two subsets.
Proposition 4.5.6 Let S and T be two subsets of the vertex set of Γ, of sizes
s and t, respectively. Let e(S, T ) be the number of ordered edges xy with x ∈ S
and y ∈ T . Then
r
kst s t √
|e(S, T ) − | ≤ λ st(1 − )(1 − ) ≤ λ st.
n n n
Proof. Write the characteristic vectors χS and χT of the sets S and PT as
a linear combination
P of a set of orthonormal eigenvectors of A: χ S =
P αi ui
and χT = βi ui √where Aui = θi u√i . Then e(S, T ) = χ⊤ S AχT = αi βi θ i .
kst
We
P have α 1 = s/ P n and β 1 = t/P n and θ 1 = k. Now |e(S, T ) − n | =
2 2
| i>1 αi βi θi | ≤ λ i>1 |αi βi | and i>1 αi ≤ (χS , χS ) − s /n = s(n − s)/n,
P p
and i>1 βi2 ≤ t(n − t)/n, so that |e(S, T ) − kst n | ≤ λ st(n − s)(n − t)/n.
If S and T are equal or complementary, this says that
kst s(n − s)
|e(S, T ) − |≤λ .
n n
In particular, the average valency kS of an induced subgraph S of size s satisfies
|kS − ks n−s
n | ≤ λ n . For example, the Hoffman-Singleton graph has θ2 = 2,
θn = −3, so λ = 3 and we find equality for subgraphs K15 (s = 15, kS = 0),
10K2 (s = 20, kS = 1) and 5C5 (s = 25, kS = 2).
4.5.4 Expansion
An expander is a graph with the property that the number of points at distance
at most one from any given (not too large) set is at least a fixed constant (larger
than one) times the size of the given set. Expanders became famous because of
their rôle in sorting networks (cf. Ajtai-Komlós-Szemerédi [1]) and have since
found many other applications. The above proposition already implies that
there cannot be too many vertices without neighbors in R. A better bound
was given by Tanner [292] (in order to show that generalized polygons are good
expanders).
Proposition 4.5.7 (Tanner [292]) Let R be a set of r vertices of Γ and let Γ(R)
be the set of vertices adjacent to some point of R. Then
|Γ(R)| ρ
≥ λ2
n ρ+ k2 (1 − ρ)
where ρ = r/n.
Proof. Let χ be the characteristic vector of R. Write P it as a linear combina-
tion of a setP of orthonormal eigenvectorsP 2 2of A: χ = αi ui where Aui = θi ui .
Then Aχ = αi θi ui and (Aχ, Aχ) = αi θi , so that kAχk2 ≤ α02 (θ02 − λ2 ) +
P 2
λ2 αi2 = (χ, u0 )2 (k 2 −λ2 )+λ2 (χ, χ) = rn (k 2 −λ2 )+rλ2 . Now let ψ be the char-
acteristic vector of Γ(R). Then k r = (Aχ, 1)2 = (Aχ, ψ)2 ≤ kAχk2 kψk2 ≤
2 2
2
|Γ(R)|.( rn (k 2 − λ2 ) + rλ2 ), proving our claim.
4.5. THE SECOND LARGEST EIGENVALUE 69
The above used two-sided bounds on the eigenvalues different from the va-
lency. It suffices to bound θ2 . Let the edge expansion constant h(Γ) of a graph Γ
be the minimum of e(S, T )/|S| where the minimum is taken over all partitions
{S, T } of the vertex set with |S| ≤ |T |, and where e(S, T ) is the number of edges
meeting both S and T . Then
p
Proposition 4.5.8 12 (k − θ2 ) ≤ h(Γ) ≤ k 2 − θ22 .
Proof. For the lower bound, apply interlacing to A and a partition {S, T }
of the vertex set, with s = |S| and t = |T |. Put e = e(S, T ). One finds
ne/st ≥ k − θ2 , so that e/s ≥ (t/n)(k − θ2 ) ≥ 12 (k − θ2 ). For the upper bound,
consider a nonnegative vector w indexed by the point set X of Γ, with support
of size at most 21 n. If wx takes t different nonzero values a1 > . . . > at > 0,
then let Si = {x | wx ≥ ai } (1 ≤ i ≤ t), and let mi = |Si \ Si−1 | (with S0 = ∅).
Let h = h(Γ). Now X X
h wx ≤ |wx − wy |.
x x∼y
Indeed, all Si have size at most 12 n, so at least h|Si | edges

P stick out of Si , and
these contribute at least h(m1 + ... + mi )(ai −P ai+1 ) to xP ∼ y |wx − wy | (with
at+1 = 0). The total contribution is at least h i mi ai = h x wx .
Let u be an eigenvector of A with Au = θ2 u. We may assume that ux > 0
for at most 21 n points x (otherwise
P replace P u by −u). Define a vector v by
vx = max(ux , 0). Since (Av)x = y ∼ x vy ≥ y ∼ x uy = (Au)x = θ2 ux = θ2 vx
P P
if vx > 0, we have v ⊤ Av = x vx (Av)x ≥ θ2 vx2 .
P P
Note that x ∼ y (vx ± vy )2 = k x vx2 ± v ⊤ Av.
2
Apply
P the above
P to the nonnegative
P vector w given P by wx = v2x .1/2We
2 2 2 2
find h x vx ≤ x ∼ y |vx − vy | ≤ ( px ∼ y (vx − vy ) . x ∼ y (vx + vy ) ) =
P 2 2 ⊤ 2 1/2
P 2 2 2
((k x vx ) − (v Av) ) ≤ ( x v x ) k − θ2 .
For similar results for not necessarily regular graphs, see §4.6.
4.5.5 Toughness and Hamiltonicity

As application of the above ideas, one can give bounds for the toughness of a
graph in terms of the eigenvalues.
A connected, noncomplete graph Γ is called t-tough if one has |S| ≥ tc
for every disconnecting set of vertices S such that the graph induced on its
complement has c ≥ 2 connected components. The toughness τ (Γ) of a graph
Γ is the largest t such that Γ is t-tough. For example, the Petersen graph has
toughness 4/3.
This concept was introduced by Chvátal [87], who hoped that t-tough graphs
would be Hamiltonian (i.e., have a circuit passing through all vertices) for suf-
ficiently large t. People tried to prove this for t = 2, the famous ‘2-tough
conjecture’, but examples were given in [18] of t-tough nonhamiltonian graphs
for all t < 9/4. Whether a larger bound on τ suffices is still open.
Still, being tough seems to help. In [19] it was shown that a t-tough graph Γ
on n ≥ 3 vertices with minimum degree δ is Hamiltonian when (t+1)(δ +1) > n.
Proposition 4.5.9 ([43]) Let Γ be a connected noncomplete regular graph of

valency k and let |θ| ≤ λ for all eigenvalues θ 6= k. Then τ (Γ) > k/λ − 2.
This proposition gives the right bound, in the sense that there are infinitely
many graphs with τ (Γ) ≤ k/λ. The constant 2 can be improved a little. The
result can be refined by separating out the smallest and the second largest
eigenvalue. The main tool in the proof is Proposition 4.5.5.
See also the remarks following Theorem 8.3.2.
Krivelevich & Sudakov [219] show that, when n is large enough, a graph
on n vertices, regular of degree k = θ1 , and with second largest eigenvalue θ2
satisfying
θ2 (log log n)2
<
θ1 1000 log n log log log n
is Hamiltonian. Pyber [261] shows that it follows that every sufficiently large
strongly regular graph is Hamiltonian.
4.5.6 The Petersen graph is not Hamiltonian

An amusing application of interlacing shows that the Petersen graph is not
Hamiltonian. Indeed, a Hamilton circuit in the Petersen graph would give an
induced C10 in its line graph. Now the line graph of the Petersen graph has
spectrum 41 25 (−1)4 (−2)5 and by interlacing the seventh eigenvalue 2 cos 53 π =
√
(1 − 5)/2 of C10 should be at most −1, contradiction. (Cf. [247, 190].)
4.5.7 Diameter bound

Chung [85] gave the following diameter bound.
Proposition 4.5.10 Let Γ be a graph on n ≥ 2 vertices, regular of valency k,
with diameter d, and with eigenvalues k = θ1 ≥ ... ≥ θn . Let λ = maxi>1 |θi |.
Then
log(n − 1)
d≤ .
log(k/λ)
Proof. The graph Γ has diameter at most m when Am > P0. Let A have
⊤
orthonormal eigenvectors ui with Aui = θi ui . Then A = θ u
i i i i u . Take
1 m
P m ⊤ km m
P
u1 = n 1. Now (A )xy = i θi (ui ui )xy ≥ n − λ
√
i>1 |(ui )x |.|(ui )y | and
P P 2 1/2
P 2 1/2
i>1 |(u )
i x |.|(u )
i y | ≤ ( i>1 |(u )
i x | ) ( i>1 |(u )
i y | ) = (1 − |(u1 )x |2 )1/2
(1 − |(u1 )y |2 )1/2 = 1 − n1 , so that (Am )xy > 0 if k m > (n − 1)λm .
4.6 Separation
Let Γ be a graph with Laplace matrix L and Laplace eigenvalues 0 = µ1 ≤
. . . ≤ µn . The Laplace matrix of a subgraph Γ′ of Γ is not a submatrix of L,
unless Γ′ is a component. So the interlacing techniques of §2.5 do not work in
such a straightforward manner here. But we can obtain results if we consider
off-diagonal submatrices of L.
Proposition 4.6.1 Let X and Y be disjoint sets of vertices of Γ, such that
there is no edge between X and Y . Then
2
|X||Y | µn − µ2
≤ .
(n − |X|)(n − |Y |) µn + µ2
4.6. SEPARATION 71
Proof. Put µ = 21 (µn + µ2 ) and define a matrix A of order 2n by

0 L − µI
A= .
L − µI 0
Let A have eigenvalues θ1 ≥ . . . ≥ θ2n . Then θ2n+1−i = −θi (1 ≤ i ≤ 2n) and

θ1 = µ and θ2 = 12 (µn − µ2 ). The sets X and Y give rise to a partitioning of A
(with rows and columns indexed by Y , Y , X, X) with quotient matrix
 
0 0 −µ 0
 0 0 |X|
−µ + µ n−|Y |X|
−µ n−|Y 
 | | 
B= |Y | |Y |  .
 −µ n−|X| −µ + µ n−|X| 0 0 
0 −µ 0 0
Let B have eigenvalues η1 ≥ . . . ≥ η4 . Then η1 = θ1 = µ and η4 = θ2n = −µ,

|X||Y |
and η1 η2 η3 η4 = det B = µ4 (n−|X|)(n−|Y |) > 0. Using interlacing we find
|X||Y |
µ2 = −η2 η3 ≤ −θ2 θ2n−1 = ( 12 (µn − µ2 ))2 ,
(n − |X|)(n − |Y |)
which gives the required inequality.

One can rewrite Tanner’s inequality (applied with R = X, Γ(R) = V Γ \ Y ) in
the form |X||Y |/(n − |X|)(n − |Y |) ≤ (λ/k)2 where λ = max(θ2 , −θn ), and this
is slightly weaker than the above, equivalent only when θn = −θ2 .
The vertex sets X and Y with the above property are sometimes called
disconnected vertex sets. In the complement X and Y become sets such that all
edges between X and Y are present. Such a pair is called a biclique.
For applications another form is sometimes handy:
Corollary 4.6.2 Let Γ be a connected graph on n vertices, and let X and Y be

disjoint sets of vertices, such that there is no edge between X and Y . Then
|X||Y | (µn − µ2 )2
≤ .
n(n − |X| − |Y |) 4µ2 µn
Proof. Let K be the constant for which Proposition 4.6.1 says |X||Y | ≤ K(n−
|X|)(n − |Y |). Then |X||Y |(1 − K) ≤ n(n − |X| − |Y |)K.
The above proposition gives bounds on vertex connectivity. For edge connec-
tivity one has
Proposition 4.6.3 (Alon & Milman [6]) Let A and B be subsets of V Γ such
that each point of A has distance at least ρ to each point of B. Let F be the set
of edges which do not have both ends in A or both in B. Then
|A||B|
|F | ≥ ρ2 µ2 .
|A| + |B|
For ρ = 1 this yields:

Corollary 4.6.4 Let Γ be a graph on n vertices, A a subset of V Γ, and F the

set of edges with one end in A and one end outside A. Then
|A|
|F | ≥ µ2 |A|(1 − ).
n
Let χ be the characteristic vector of A. Then equality holds if and only if χ− |A|
n 1
is a Laplace eigenvector with eigenvalue µ2 .
Proof. Let ui be an orthonormal systemP of Laplace eigenvectors, so that

P Lu i =
µi ui . Take u1 = √1n 1. Let χ = αi ui . Now |A| = (χ, χ) = αi2 and
P P
α1 = (χ, u1 ) = √1n |A|. We find |F | = a∈A,b6∈A,a ∼ b 1 = x ∼ y (χx − χy )2 =
P 2 P
χ⊤ Lχ = αi µi ≥ ( i>1 αi2 )µ2 .
This is best possible in many situations.
Example The Hoffman-Singleton graph has Laplace spectrum 01 528 1021 and
we find |F | ≥ |A||B|/10 and this holds with equality for the 10-40 split into a
Petersen subgraph and its complement.
4.6.1 Bandwidth
A direct consequence of Proposition 4.6.1 is an inequality of Helmberg, Mohar,
Poljak and Rendl [189], concerning the bandwidth of a graph. A symmetric
matrix M is said to have bandwidth w if (M )i,j = 0 for all i, j satisfying
|i−j| > w. The bandwidth w(Γ) of a graph Γ is the smallest possible bandwidth
for its adjacency matrix (or Laplace matrix). This number (and the vertex order
realizing it) is of interest for some combinatorial optimization problems.
l m
Theorem 4.6.5 Suppose Γ is not edgeless and define b = n µµn2 , then

b if n − b is even,
w(Γ) ≥
b−1 if n − b is odd.
Proof. Order the vertices of Γ such that L has bandwidth w = w(Γ). If n − w

is even, let X be the first 12 (n − w) vertices and let Y be the last 21 (n − w)
vertices. Then Proposition 4.6.1 applies and thus we find the first inequality. If
n − w is odd, take for X and Y the first and last 12 (n − w − 1) vertices and the
second inequality follows. If b and w have different parity, then w − b ≥ 1 and
so the better inequality holds.
In case n − w is odd, the bound can be improved a little by applying Proposi-
tion 4.6.1 with |X| = 12 (n − w + 1) and |Y | = 21 (n − w − 1). It is clear that the
result remains valid if we consider graphs with weighted edges.
4.6.2 Perfect matchings

A more recent application of Proposition 4.6.1 is the following sufficient con-
dition for existence of a perfect matching (a perfect matching in a graph is a
subset of the edges, such that every vertex of the graph is incident with exactly
one edge of the subset).
4.6. SEPARATION 73
Theorem 4.6.6 ([53]) Let Γ be a graph with n vertices, and Laplacian eigen-
values 0 = µ1 ≤ µ2 ≤ . . . ≤ µn . If n is even and µn ≤ 2µ2 , then Γ has a perfect
matching.
Except for Proposition 4.6.1, we need two more tools. The first one is Tutte’s
famous characterization of graphs with a perfect matching. The second one is
an elementary observation.
Theorem 4.6.7 (Tutte [296]) A graph Γ = (V, E) has no perfect matching if

and only if there exists a subset S ⊂ V , such that the subgraph of Γ induced by
V \ S has more than |S| odd components.
Pn
Lemma 4.6.8 Let x1 . . . xn be n positive integers such that i=1 xi = k ≤ 2n−
P every integer ℓ, satisfying 0 ≤ ℓ ≤ k, there exists an I ⊂ {1, . . . , n}
1. Then for
such that i∈I xi = ℓ.
Proof. Induction on n. The case n = 1 is trivial. If n ≥ 2, assume x1 ≥ . . . ≥

xn .PThen n − 1 ≤ k − x1 ≤ 2(n − 1) − 1 and we apply the induction hypothesis
n
to i=2 xi = k − x1 with the same ℓ if ℓ ≤ n − 1, and ℓ − x1 otherwise.
Proof of Theorem 4.6.6. Assume Γ = (V, E) has no perfect matching. By
Tutte’s theorem there exists a set S ⊂ V of size s (say), such that the subgraph
Γ′ of Γ induced by V \ S has q > s odd components. But since n is even, s + q
is even, hence q ≥ s + 2.
First assume n ≤ 3s + 3. Then Γ′ has at most 2s + 3 vertices and at least s + 2
components. By Lemma 4.6.8, Γ′ and hence Γ, has a pair of disconnected vertex
sets X and Y with |X| = ⌊ 12 (n−s)⌋ and |Y | = ⌈ 12 (n−s)⌉. Now Proposition 4.6.1
implies
2
µn − µ2 |X| · |Y | (n − s)2 − ǫ
≥ = ,
µn + µ2 ns + |X| · |Y | (n + s)2 − ǫ
where ǫ = 0 if n − s is even and ǫ = 1 if n − s is odd. Using n ≥ 2s + 2 we obtain
µn − µ2 n−s−1 s+1 1
> ≥ > .
µn + µ2 n+s 3s + 2 3
Hence 2µ2 < µn .
Next assume n ≥ 3s + 4. Now Γ′ , and hence Γ, has a pair of disconnected
vertex sets X and Y with |X| + |Y | = n − s and min{|X|, |Y |} ≥ s + 1, so
|X| · |Y | ≥ (s + 1)(n − 2s − 1) > ns − 2s2 . Now Proposition 4.6.1 implies
2
µn − µ2 |X| · |Y | ns − 2s2 1 s 1
≥ ≥ 2
= − > ,
µn + µ2 ns + |X| · |Y | 2ns − 2s 2 2n − 2s 4
by use of n ≥ 3s + 4. So
µn − µ2 1 1
> > ,
µn + µ2 2 3
hence 2µ2 < µn .
The complete bipartite graphs Kl,m with l ≤ m have Laplacian eigenvalues
µ2 = m and µn = n = l + m. This shows that 2µ2 can get arbitrarily close to
µn for graphs with n even and no perfect matching.
If the graph is regular, the result can be improved considerably.
Theorem 4.6.9 [53, 89] A connected k-regular graph on n vertices, where n is

even, with (ordinary) eigenvalues k = λ1 ≥ λ2 . . . ≥ λn , which satisfies
( 3
k − 1 + k+1 if k is even,
λ3 ≤ 4
k − 1 + k+2 if k is odd,
has a perfect matching.
Proof. Let Γ = (V, E) be a k-regular graph with n = |V | even and no perfect

matching. By Tutte’s Theorem 4.6.7 there exists a set S ⊂ V of size s such that
V \ S induces a subgraph with q ≥ s + 2 odd components Γ1 , Γ2 , . . . , Γq (say).
Let ti denote the number of edges in Γ Pbetween S and Γi , and let ni be the
q
number of vertices of Γi . Then clearly i=1 ti ≤ ks, s ≥ 1, and ti ≥ 1 (since Γ
is connected). Hence ti < k and ni > 1 for at least three values of i, say i = 1,
2 and 3. Let ℓi denote the largest eigenvalue of Γi , and assume ℓ1 ≥ ℓ2 ≥ ℓ3 .
Then eigenvalue interlacing applied to the subgraph induced by the union of Γ1 ,
Γ2 and Γ3 gives ℓi ≤ λi for i = 1, 2, 3.
Consider Γ3 with n3 vertices and e3 edges (say). Then 2e3 = kn3 − t3 ≤
n3 (n3 −1). We saw that t3 < k and n3 > 1, hence k < n3 . Moreover, the average
degree d3 of Γ3 equals 2e3 /n3 = k − t3 /n3 . Because n3 is odd and kn3 − t3 is
even, k and t3 have the same parity, therefore t3 < k implies t3 ≤ k − 2. Also
k < n3 implies k ≤ n3 − 1 if k is even, and k ≤ n3 − 2 if k is odd. Hence
(
k − k−2
k+1 if k is even,
d3 ≥ k−2
k − k+2 if k is odd.
Note that t3 < n3 implies that Γ3 cannot be regular. Next we use the fact
that the largest adjacency eigenvalue of a graph is bounded from below by the
average degree with equality if and only if the graph is regular (Proposition
3.1.2). Thus d3 < ℓ3 . We saw that ℓ3 ≤ λ3 , which finishes the proof.
From the above it is clear that n even and λ2 ≤ k − 1 implies existence of a
perfect matching. In terms of the Laplacian matrix this translates into:
Corollary 4.6.10 A regular graph with an even number of vertices and alge-
braic connectivity at least 1 has a perfect matching.
But we can say more. The Laplacian matrix of a disjoint union of n/2 edges
has eigenvalues 0 and 2. This implies that deletion of the edges of a perfect
matching of a graph Γ reduces the eigenvalues of the Laplacian matrix of Γ by
at most 2 (by the Courant-Weyl inequalities 2.8.1). Hence:
Corollary 4.6.11 A regular graph with an even number of vertices and alge-
braic connectivity µ2 has at least ⌊(µ2 + 1)/2⌋ disjoint perfect matchings.
Cioabă, Gregory and Haemers [90] have improved the sufficient condition for
a perfect matching from Theorem 4.6.9 to λ3 <√ϑk where ϑ3 = 2.85577... (the
largest root of x3 −p
x2 − 6x + 2), ϑk = (k − 2 + k 2 + 12)/2 if k ≥ 4 and even,
and ϑk = (k − 3 + (k + 1)2 + 16)/2 if k ≥ 5 and odd. They also prove that
this bound is best possible by giving examples of k-regular graphs with n even,
and λ3 = ϑk that have no perfect matching. The example for k = 3 is presented
in Figure 4.1.
4.7. BLOCK DESIGNS 75
tH tH tH

t
H H
t
H
t H
t
H
t H
t
\ Ht \ Ht \ Ht
\\t \\t \\
HH
t
HH
HHt
Figure 4.1: A 3-regular graph with no perfect matching
4.7 Block Designs

In case we have a non-symmetric matrix N (say) we can still use interlacing by
considering the matrix
0 N
A=
N⊤ 0
like we did in the previous section. Then we find results in terms of the eigen-
values of A, which now satisfy θi = −θn−i+1 for i = 1, . . . , n. The positive
eigenvalues of A are the singular values of N , they are also the square roots of
the non-zero eigenvalues of N N ⊤ (and of N ⊤ N ).
Suppose N is the 0-1 incidence matrix of an incidence structure (P, B) with
point set P (rows) and block set B (columns). Then we consider the so-called
incidence graph Γ of (P, B), which is the bipartite graph with vertex set P ∪ B,
where two vertices are adjacent if they correspond to an incident point-block
pair. An edge of Γ is called a flag of (P, B).
An incidence structure (P, B) is called a t-(v, k, λ) design if |P | = v, all blocks
are incident with k points, and for every t-set of points there are exactly λ blocks
incident with all t points. For example (P, B) is a 1-(v, k, r) design precisely
when N has constant column sums k (i.e. N ⊤ 1 = k1), and constant row sums
r (i.e. N 1 = r1), in other words Γ is biregular with degrees k and r. Moreover,
(P, B) is a 2-(v, k, λ) design if and only if N ⊤ 1 = k1 and N N ⊤ = λJ + (r − λ)I.
Note that for t ≥ 1, a t-design is also a (t − 1)-design. In particular a 2-(v, k, λ)
design is also a 1-(v, k, r) design with r = λ(v − 1)/(k − 1).
Theorem 4.7.1 Let (P, B) be a 1-(v, k, r) design with b blocks and let (P ′ , B ′ )
be a substructure with m′ flags. Define b = |B|, v ′ = |P ′ | and b′ = |B ′ |. Then
v b
(m′ − b′ k)(m′ ′ − v ′ r) ≤ θ22 (v − v ′ )(b − b′ ) .
v′ b
Equality implies that all four substructures induced by P ′ or V \ V ′ and B ′ or
B \ B ′ form a 1-design (possibly degenerate).
Proof. We apply Corollary 2.5.4. The substructure (P ′ , B ′ ) gives rise to a

partition of A with the following quotient matrix
 
m′ m′
0 0 v ′ r − v ′
 0 0 b′ k−m′ ′
k−m′ 
r − b v−v
 v−v ′ ′ 
B =  m′ m′  .
 b′ ′ k − b′ 0 0 
′ ′ ′
v r−m v r−m
b−b′ k − b−b′ 0 0
√
We easily have θ1 = −θn = η1 = −η4 = rk and
!
m′ vv′ − b′ k m′ bb′ − v ′ r
det(B) = rk .
v − v′ b − b′
Interlacing gives
det(B)
= −η2 η3 ≤ −θ2 θn−1 = θ22 ,
rk
which proves the first statement. If equality holds then θ1 = η1 , θ2 = η2 ,
θn−1 = η3 and θn = η4 , so we have tight interlacing, which implies the second
statement.
The above result becomes especially useful if we can express θ2 in terms of
the design parameters. For instance if (P, B) is a 2-(v, k, λ) design, then θ22 =
r − λ = λ v−k
k−1 (see exercises) and if (P, B) is a generalized quadrangle of order
(s, t), then λ22 = s + t (see §8.6). Let us consider two special cases. (A 2-design
(P, B) with |P | = |B| is called symmetric.)
Corollary 4.7.2 If a symmetric 2-(v, k, λ) design (P, B) has a symmetric 2-

(v ′ , k ′ , λ′ ) subdesign (P ′ , B ′ ) (possibly degenerate) then
2
(k ′ v − kv ′ )2 ≤ (k − λ)(v − v ′ ) .
If equality holds, then the subdesign (P ′ , B \ B ′ ) is a 2-(v ′ , v ′ (k − k ′ )/(v − v ′ ), λ −

λ′ ) design (possibly degenerate).
Proof. In Theorem 4.7.1 take b = v, r = k, b′ = v ′ , m′ = v ′ k ′ and θ22 = k − λ.

Corollary 4.7.3 Let X be a subset of the points and let Y be a subset of the
blocks of a 2-(v, k, λ) design (P, B), such that no point of X is incident with a
block of Y , then
kr|X||Y | ≤ (r − λ)(v − |X|)(b − |Y |).
If equality holds then the substructure (X, B ′ ) = (X, B \ Y ) is a 2-design.
Proof. Take m′ = 0, v ′ = |X|, b′ = |Y | and θ22 = r − λ. Now 4.7.1 gives

the inequality and that (X, B ′ ) is a 1-design. But then (X, B ′ ) is a 2-design,
because (P, B) is.
An example of a subdesign of a symmetric design is the incidence structure
formed by the absolute point and lines of a polarity in a projective plane of
order q. This gives a (degenerate) 2-(v ′ , 1, 0) design in a 2-(q 2 + q + 1, q + 1, 1)
√
design. The bound gives v ′ ≤ q q + 1. (See also the following section.) The 2-
√
(q q + 1, q + 1, 1) design, which is obtained in case of equality is called a unital.
Other examples of symmetric designs that meet the bound can be found in
Haemers & Shrikhande [182] or Jungnickel [210]. Wilbrink used Theorem 4.7.1
to shorten the proof of Feit’s result on the number of points and blocks fixed
by an automorphism group of a symmetric design (see [61]). The inequality of
the second corollary is for example tight for hyperovals and (more generally)
maximal arcs in finite projective planes.
4.8. POLARITIES 77
4.8 Polarities
A projective plane is a point-line geometry such that any two points are on a
unique line, and any two lines meet in a unique point. It is said to be of order
q when all lines have q + 1 points and all points are on q + 1 lines. A projective
plane of order q has q 2 + q + 1 points and as many lines.
A polarity of a point-block incidence structure is a map of order 2 inter-
changing points and blocks and preserving incidence. An absolute point is a
point incident with its image under the polarity.
Suppose we have a projective plane of order q with a polarity σ. The polarity
enables us to write the point-line incidence matrix N as a symmetric matrix, and
then the number of absolute points is tr N . By definition we have N 2 = N N ⊤ =
J +qI, which has one eigenvalue equal to (q +1)2 and all other eigenvalues equal
√ √
to q. That means that N has spectrum (q+1)1 , q m , − q n , for certain integers
m, n, where this time exponents indicate multiplicities. The number of absolute
√
points equals a = q + 1 + (m − n) q. It follows that if q is not a square then
m = n and there are precisely q + 1 absolute points. If q is a square, and p is a
prime dividing q, then a ≡ 1 (mod p) so that a is nonzero.
(This is false in the infinite case: the polarity sending the point (p, q, r) to
the line pX + qY + rZ = 0 has no absolute points over R.)
With slightly more effort one finds bounds for the number of absolute points:
Proposition 4.8.1 A polarity of a projective plane of order q has at least q + 1

√
and at most q q + 1 absolute points.
Proof. Suppose z is a non-absolute point. Now σ induces a map τ on the line

z σ defined for y ∈ z σ by: y τ is the common point of y σ and z σ . Now τ 2 = 1, and
y τ = y precisely when y is absolute. This shows that the number of absolute
points on a non-absolute line is q + 1 (mod 2).
Now if q is odd, then take an absolute point x. This observation says that
each line on x different from xσ contains another absolute point, for a total of at
least q + 1. On the other hand, if q is even, then each non-absolute line contains
an absolute point, so that q 2 + q + 1 − a ≤ aq and a ≥ q + 1.
For the upper bound, use interlacing: partition the matrix N into ab-
solute
/ non-absolute points/lines and find the matrix of average row sums
1 q aq
aq aq where v = q 2 + q + 1, with eigenvalues q + 1 and 1 − v−a .
v−a q + 1 − v−a
aq √ √
Now interlacing yields 1 − v−a ≥ − q, that is, a ≤ q q + 1, just like we found
in the previous section.
The essential part of the proof of the lower bound was to show that there is at
least one absolute point, and this used an eigenvalue argument.
4.9 Exercises
Exercise 1 Deduce Proposition 4.2.3(iii) (the part that says (m − 1)θt+1 +
θn−t(m−1) ≥ 0) from Theorem 4.1.3.
Exercise 2 An (ℓ, m)-biclique in a graph Γ is a complete bipartite subgraph
Kℓ,m of Γ (not necessarily induced). Let 0 = µ1 ≤ . . . ≤ µn be the Laplace
eigenvalues of Γ. Show that ℓm/(n − ℓ)(n − m) ≤ ((µn − µ2 )/(2n − µ2 − µn ))2

if Γ is non-complete and contains an (ℓ, m)-biclique.
Exercise 3 Let A be the incidence graph of a 2-(v, k, λ) design with b blocks
and r blocks incident with each point. Express the spectrum of A in the design
parameters v, k, λ, b and r.
Exercise 4 Let (P, B) is a 2-(v, k, λ) design, and suppose that some block
is repeated ℓ times (i.e. ℓ blocks are incident with exactly the same set of k
points). Prove that b ≥ ℓv (this is Mann’s inequality).
Chapter 5
Trees
Trees have a simpler structure than general graphs, and we can prove stronger
results. For example, interlacing tells us that the multiplicity of an eigenvalue
decreases by at most one when a vertex is removed. For trees Godsil’s Lemma
gives the same conclusion also when a path is removed.
5.1 Characteristic polynomials of trees

For a graph Γ with adjacency matrix A, let φΓ (t) := det(tI − A) be its charac-
teristic polynomial.
Note that since the characteristic polynomial of the disjoint union of two
graphs is the product of their characteristic polynomials, results for trees im-
mediately yield results for forests as well.
It will be useful to agree that φT \x,y = 0 if x = y.
Proposition 5.1.1 Let T be a tree, and for x, y ∈ T , let Pxy be the unique path
joining x and y in T .
(i) Let e = xy be an edge in T that separates T into two subtrees A and B,
with x ∈ A and y ∈ B. Then
φT = φA φB − φA\x φB\y .
(ii) Let x be a vertex of T . Then

X
φT (t) = tφT \x (t) − φT \{x,y} (t).
y∼x
(iii) Let x be a vertex of T . Then

X
φT \x (t)φT (s) − φT \x (s)φT (t) = (s − t) φT \Pxy (s)φT \Pxy (t).
y∈T
(iv) Let x be a vertex of T . Then

X
φT \x φ′ T − φ′ T \x φT = φ2T \Pxy .
y∈T
79
80 CHAPTER 5. TREES
(v) Let x, y be vertices of T . Then
φT \x φT \y − φT \x,y φT = φ2T \Pxy .
(vi) Let x, y, z be vertices of T where z ∈ Pxy . Then
φT \x,y,z φT = φT \x φT \y,z − φT \z φT \x,y + φT \y φT \x,z .

P
(vii) We have φ′ T = x∈T φT \x .
(viii) Let T have n vertices and cm matchings of size m. Then
X
φT (t) = (−1)m cm tn−2m .
m
Proof. Part (i) follows by expansion of the defining determinant. It can also
be phrased as φT = φT \e − φT \{x,y} . Part (ii) follows by applying (i) to all edges
on x. Note that φ{x} (t) = t. Part (iii) follows from (ii) by induction on the size
of T : expand in the LHS φT (s) and φT (t) using (ii), and then use induction.
Part (iv) is immediate from (iii). Part (vii) follows by taking the derivative
of the defining determinant. Part (viii) is a reformulation of the description
in §1.2.1. Note that the only directed cycles in a tree are those of length 2.
Part (v) is true if T = Pxy , and the general case follows from part (vi) and
induction: the statement remains true when a subtree S is attached via an edge
e at a vertex z ∈ Pxy . Finally, part (vi) follows from: if Γ \ z = A + B, then
φΓ = φA∪z φB + φA φB∪z − φA φ{z} φB , where of course φ{z} (t) = t.
Theorem 5.1.2 (‘Godsil’s Lemma’, [157]) Let T be a tree and θ an eigenvalue

of multiplicity m > 1. Let P be a path in T . Then θ is eigenvalue of T \ P with
multiplicity at least m − 1.
Proof. By parts (iv) and (vii) of the above Proposition we have

X
φ′ T (t)2 − φ′′ T (t)φT (t) = φT \Pxy (t)2 .
x,y∈T
Now θ is a root of multiplicity at least 2m − 2 of the left hand side, and hence
also of each of the terms on the right hand side.
As an application of Godsil’s Lemma, consider a tree T with e distinct eigen-
values and maximum possible diameter e − 1. Let P be a path of length e − 1
(that is, with e vertices) in T . Then T \ P has a spectrum that is independent
of the choice of P : for each eigenvalue θ with multiplicity m of T , the forest
T \ P has eigenvalue θ with multiplicity m − 1 (and it has no other eigenvalues).
In particular, all eigenvalues of a path have multiplicity 1.
Note that going from T to T \ x changes multiplicities by at most 1: they
go up or down by at most one. Godsil’s Lemma is one-sided: going from T to
T \ P , the multiplicities go down by at most one, but they may well go up by
more. For example, if one joins the centers x, y of two copies of K1,m by an
edge, one obtains a tree T that has 0 as an eigenvalue of multiplicity 2m − 2.
For P = xy the forest T \ P has 0 with multiplicity 2m.
5.2. EIGENVECTORS AND MULTIPLICITIES 81
5.2 Eigenvectors and multiplicities

For trees we have rather precise information about eigenvectors and eigenvalue
multiplicities (Fiedler [142]).
Lemma 5.2.1 Let T be a tree with eigenvalue θ, and let Z = ZT (θ) be the set
of vertices in T where all θ-eigenvectors vanish. If for some vertex t ∈ T some
component S of T \ t has eigenvalue θ (in particular, if some θ-eigenvector of
T vanishes at t), then Z 6= ∅.
Proof. Consider proper subtrees S of T with eigenvalue θ and with a single

edge st joining some vertex s ∈ S with some vertex t ∈ T \S, and pick a minimal
one. If |S| = 1, then θ = 0, and t ∈ Z. Assume |S| > 1. If a θ-eigenvector
u of S is the restriction to S of a θ-eigenvector v of T , then v vanishes in t.
So, if some θ-eigenvector v of T does not vanish at t then u and v|S are not
dependent, and some linear combination vanishes in s and is a θ-eigenvector of
S \ s, contradicting minimality of S. This shows that t ∈ Z.
Note that it is not true that the hypothesis of the lemma implies that t ∈ Z.
For example, consider the tree T of type D6 given by 1 ∼ 2 ∼ 3 ∼ 4 ∼ 5, 6.
It has Z(0) = {2, 4}, and the component S = {4, 5, 6} of T \ 3 has eigenvalue 0,
but 3 6∈ Z(0).
Proposition 5.2.2 Consider a tree T with eigenvalue θ, and let Z = Z(θ) be

the set of vertices in T where all θ-eigenvectors vanish. Let Z0 = Z0 (θ) be the
set of vertices in Z that have a neighbor in T \ Z.
(i) Let S be a connected component of T \ Z. Then S has eigenvalue θ with
multiplicity 1. If u is a θ-eigenvector of S, then u is nowhere zero.
(ii) Let T \ Z have c connected components, and let d = |Z0 |. Then θ has
multiplicity c − d.
The components of T \ Z(θ) are called the eigenvalue components of T for θ.
Proof. (i) Suppose θ is eigenvalue of T with multiplicity greater than 1. Then

some eigenvector has a zero coordinate and hence induces a θ-eigenvector on a
proper subtree.
By Lemma 5.2.1, Z is nonempty.
If S is a connected component of T \ Z then it has eigenvalue θ (otherwise
S ⊂ Z, contradiction). Apply Lemma 5.2.1 to S instead of T to find that if
some θ-eigenvector of S vanishes on a point of S, then there is a point s ∈ S
where all of its θ-eigenvectors vanish. But the restriction to S of a θ-eigenvector
of T is a θ-eigenvector of S, so s ∈ Z, contradiction.
(ii) Each point of Z0 imposes a linear condition, and since T is a tree, these
conditions are independent.
We see that if the multiplicity of θ is not 1, then Z contains a vertex of

degree at least three. In particular, Z 6= ∅, and hence Z0 6= ∅. Deleting a vertex
in Z0 from T increases the multiplicity of θ.
As an application we see that all eigenvalues of a path have multiplicity 1.

82 CHAPTER 5. TREES
5.3 Sign patterns of eigenvectors of graphs

For a path, the i-th largest eigenvalue has multiplicity 1 and an eigenvector with
i − 1 sign changes, that is, i areas of constant sign. It is possible to generalize
this observation to more general graphs.
Given a real vector u, let the support supp u be the set {i|ui 6= 0}. For ∗ one
of <, >, ≤, ≥ we also write supp∗ u for {i|ui ∗ 0}. Let N (u) (resp. N ∗ (u)) be the
number of connected components C of the subgraph induced by supp u (resp.
supp∗ u) such that u does not vanish identically on C. Let Nθ (u) be the number
of connected components C of the subgraph induced by supp u such that u (does
not vanish identically on C and) induces an eigenvector with eigenvalue θ on C.
Proposition 5.3.1 Let Γ be a graph with eigenvalues θ1 ≥ ... ≥ θn , and let

θ = θj = θj+m−1 be an eigenvalue with multiplicity m. Let u be a vector with
Au ≥ θu. Let ∆ be the subgraph of Γ induced by supp u, with eigenvalues
η1 ≥ ... ≥ ηt . Then
(i)
N > (u) + N < (u) ≤ #{i | ηi ≥ θ} ≤ j + m − 1
(ii)
N > (u) + N < (u) − Nθ (u) ≤ #{i | ηi > θ} ≤ j − 1
(iii) if Γ has c connected components, then
N ≥ (u) + N ≤ (u) ≤ j + c − 1.
Proof. For a subset S of the vertex set of Γ, let IS be the diagonal matrix
with ones on the positions indexed by elements of S and zeros elsewhere.
Let C run through the connected components of supp> u and supp (u) + N < (u) (resp. N ≥ (u) + N ≤ (u)).
Let A be the adjacency matrix of ∆ (resp. Γ). Define a real symmetric
matrix B by BCD = u⊤ C (A − θI)uD . Then B has nonnegative row sums and
nonpositive off-diagonal entries, so B is positive semidefinite. It follows that
for y ∈ U we have y ⊤ (A − θI)y ≥ 0. This means that U intersects the space
spanned by the eigenvectors of A − θI with negative eigenvalue in 0.
For (i) N > (u) + N < (u) ≤ #{i | ηi ≥ θ} follows.
The vectors y ∈ U with y ⊤ (A − θI)y = 0 correspond to eigenvectors with
eigenvalue 0 of B, and by Lemma 2.10.1 there are at most Nθ (u) (resp. c)
independent such. This proves (ii) (resp. (iii)).
Remark For j = 1 the results follow from Perron-Frobenius. (If Γ is connected,
then the eigenvector for θ1 is nowhere zero and has constant sign.)
Examples a) Let Γ be connected and bipartite, and let θ be the smallest
eigenvalue of Γ. The corresponding eigenvector u has different signs on the
two sides of the bipartition, so supp> u and supp (u) + N < (u) = n and N (u) = 1. We have equality in (i)–(iii).
√ 1 √
b) Let Γ be the star K1,s . The spectrum is s , 0s−1 , (− s)1 . Let u be
an eigenvector with eigenvalue θ = 0 that has t nonzero coordinates. (Then
2 ≤ t ≤ s.) Now N > (u) + N < (u) = N (u) = t and N ≥ (u) + N ≤ (u) = 2, and for
t = s equality holds in (i)–(iii).
5.4. SIGN PATTERNS OF EIGENVECTORS OF TREES 83
c) Let Γ be the Petersen graph. It has spectrum 31 , 15 , (−2)4 . Let u be

an eigenvector with eigenvalue θ = 1 that vanishes on 4 points, so that supp u
induces 3K2 with spectrum 13 , (−1)3 . We find N > (u) + N < (u) = N (u) = 3
and N ≥ (u) + N ≤ (u) = 2, again equality in (i)–(iii).
d) Let Γ be the path Pn . The eigenvalues are θk = 2 cos(kπ/(n + 1)) for
k = 1, ..., n. The eigenvector u corresponding to θk has k − 1 sign changes, so
that N > (u) + N < (u) = k. If gcd(k, n + 1) = 1 then u has no zero entries, so
that N (u) = 1. Now we have equality in (i)–(iii). If gcd(k, n + 1) = r, then u
has r − 1 zero entries, so that N (u) = r. Also, the eigenvalue θk is the k/r-th of
each component of supp u, so that #{i | ηi ≥ θ} = k and #{i | ηi > θ} = k − r,
with equality in (i) and the first inequality of (ii).
Remark It is not true that N (u) ≤ m if m is the multiplicity of θ for Γ. For

example, in case b) above we have N (u) = s and m = s − 1. (And in case c)
the opposite happens: N (u) = 3 and m = 5.)
5.4 Sign patterns of eigenvectors of trees

Proposition 5.4.1 Let T be a tree with eigenvalue θ, and put Z = Z(θ). Let
T \ Z have eigenvalues η1 ≥ ... ≥ ηm . Let g = #{i | ηi ≥ θ} and h =
#{i | ηi > θ}. Let u be a θ-eigenvector of T . Then N > (u) + N < (u) = g and
N > (u) + N < (u) − N (u) = h.
Proof. Since N () and g and h are additive over connected components, we

may assume that Z is empty. Now by Proposition 5.2.2(i), θ has multiplicity
1 and u is nowhere 0. Let T have n vertices, and let there be p edges xy with
ux uy > 0 and q edges xy with ux uy < 0. Then p + q = n − 1. Since T is
bipartite, also −θ is an eigenvalue, and an eigenvector v for −θ is obtained by
switching the sign of u on one bipartite class. By Proposition 5.3.1 we have
q = N > (u) + N < (u) − 1 ≤ h and p = N > (v) + N < (v) − 1 ≤ n − h − 1, that is
q ≥ h, and hence equality holds everywhere.
Let a sign change for an eigenvector u of T be an edge e = xy such that
ux uy < 0.
Proposition 5.4.2 Let T be a tree with j-th eigenvalue θ. If u is an eigenvector

for θ with s sign changes, and d = |Z0 (θ)|, then d + s ≤ j − 1
Proof. Let T \ Z have c connected components, and let u be identically zero

on c0 of these. Then s + c − c0 = N > (u) + N < (u). Let θ = θj = θj+m−1 ,
where m = c − d is the multiplicity of θ. By Proposition 5.3.1(i) we have
s + c − c0 ≤ j + m − 1, that is, d + s − c0 ≤ j − 1. But we can make c0 zero by
adding a small multiple of some eigenvector that is nonzero on all of T \ Z.
Example For T = E6 all eigenvalues have multiplicity 1, and N > (u) + N < (u)
takes the values 1, 2, 3, 4, 4, 6 for the six eigenvectors u. The sign patterns are:
•+ •0 •−
• • • • • • • • • • • • • • •
+ + + + + + + 0 − − + + − + +
84 CHAPTER 5. TREES
•+ •0 •−
• • • • • • • • • • • • • • •
+ − − − + + − 0 + − + − + − +
We see that a small perturbation that would make u nonzero everywhere would
give the two zeros in the second eigenvector the same sign, but the two zeros
in the fifth eigenvector different sign (because θ2 > 0 and θ5 < 0) and for the
perturbed vector u′ we would find 0, 1, 2, 3, 4, 5 sign changes.
5.5 The spectral center of a tree

There are various combinatorial concepts ‘center’ for trees. One has the cen-
ter/bicenter and the centroid/bicentroid. Here we define a concept of center
using spectral methods. Closely related results can be found in Neumaier [251].
Proposition 5.5.1 Let T be a tree (with at least two vertices) with second
largest eigenvalue λ. Then there is a unique minimal subtree Y of T such
that no connected component of T \ Y has largest eigenvalue larger than λ. If
Z(λ) 6= ∅ (and in particular if λ has multiplicity larger than 1) then Y = Z0 (λ)
and |Y | = 1. Otherwise |Y | = 2, and Y contains the endpoints of the edge on
which the unique λ-eigenvector changes sign. In this latter case all connected
components of T \ Y have largest eigenvalue strictly smaller than λ.
We call the set Y the spectral center of T .
Proof. If for some vertex y all connected components of T \ y have largest
eigenvalue at most λ, then pick Y = {y}. Otherwise, for each vertex y of T
there is a unique neighbor y ′ in the unique component of T \ y that has largest
eigenvalue more than λ. Since T is finite, we must have y ′′ = y for some vertex
y. Now pick Y = {y, y ′ }. Clearly Y has the stated property and is minimal.
If Z = ∅ then λ has multiplicity 1 and by Proposition 5.4.2 there is a unique
edge e = pq such that the unique λ-eigenvector has different signs on p and q,
and both components of T \ e have largest eigenvalue strictly larger than λ, so
that Y must contain both endpoints of e.
If Z 6= ∅, then all eigenvalue components for λ have eigenvalue λ, and any
strictly larger subgraph has a strictly larger eigenvalue, so Y must contain Z0 .
By Proposition 5.4.2 we have |Z0 | = 1, say Z0 = {y}. If Y is not equal to {y},
then Y also contains y ′ . This proves uniqueness.
Suppose that Z0 = {y} and T \ y has a component with eigenvalue larger
than λ. Let u be a vector that is 0 on y, and induces an eigenvector with the
largest eigenvalue on each component of T \y. Let c be the number of connected
components of T \ Z. Proposition 5.3.1(i) now gives c + 1 ≤ 2 + (c − 1) − 1, a
contradiction. This shows that |Y | = 1 when Z is nonempty.
Finally, suppose that Y = {y, y ′ } and that T \ Y has largest eigenvalue λ.
By Lemma 5.2.1 Z 6= ∅, contradiction.
Example If T is the path Pn with n vertices, then λ = 2 cos 2π/(n + 1). If

n = 2m + 1 is odd, then Y consists of the middle vertex, and T \ Y is the union
of two paths Pm , with largest eigenvalue λ = 2 cos π/(m + 1). If n = 2m is
even, then Y consists of the middle two vertices, and T \ Y is the union of two
paths Pm−1 , with largest eigenvalue 2 cos π/m < λ.
5.6. INTEGRAL TREES 85
5.6 Integral trees

An integral tree is a tree with only integral eigenvalues. Such trees are rare. A
list of all integral trees on at most 50 vertices can be found in [46].
A funny result is
Proposition 5.6.1 (Watanabe [304]) An integral tree cannot have a perfect

matching, that is, must have an eigenvalue 0, unless it is K2 .
Proof. The constant term of the characteristic polynomial of a tree is, up to

sign, the number of perfect matchings. It is also the product of all eigenvalues.
If it is nonzero, then it is 1, since the union of two distinct perfect matchings
contains a cycle. But then all eigenvalues are ±1 and P3 is not an induced
subgraph, so we have K2 .
This result can be extended a little. Let SK1,m be the tree on 2m +√1 ver-
tices obtained by subdividing all edges of K1,m . The spectrum is ± m + 1
(±1)m−1 0.
Proposition 5.6.2 (Brouwer [46]) If an integral tree has eigenvalue 0 with mul-
tiplicity 1, then it is SK1,m , where m = t2 − 1 for some integer t ≥ 1.
For a long time it has been an open question whether there exist integral
trees of arbitrarily large diameter. Recently, this was settled in the affirmative
by Csikvári. The construction is as follows. Define trees T ′ (r1 , . . . , rm ) by
induction: T ′ () is the tree with a single vertex x0 . T ′ (r1 , . . . , rm ) is the tree
obtained from T ′ (r1 , . . . , rm−1 ) by adding rm pendant edges to each vertex u
with d(u, x0 ) = m − 1 (mod 2). The diameter of this tree is 2m (assuming
r1 > 1) and it has 2m + 1 distinct eigenvalues:
Proposition 5.6.3 (Csikvári [102]) The tree T ′ (r1 , . . . , rm ) has eigenvalues 0

√
and ± si (1 ≤ i ≤ m), where si = ri + · · · + rm .
Now all trees T ′ (n12 − n22 , . . . , nm−1

2 2
− nm 2
, nm ) are integral of diameter 2m
when n1 > n2 > . . . > nm .
A short proof can be given using the following observation. If A and B are
trees with fixed vertices x and y, respectively, then let A ∼ mB be the tree
constructed on the union of A and m copies of B, where x is joined to the m
copies of y. Now Proposition 5.1.1(i) and induction immediately yields that
m−1
T = A ∼ mB has characteristic polynomial φT = φB (φA φB − mφA\x φB\y ),
where the last factor is symmetric in A and B.
Proof. Induction on m. The statement holds for m ≤ 1. With A = T ′ (r3 , . . .)
and B = T ′ (r2 , r3 , . . .) we have T ′ (r1 , r2 , r3 , . . .) = A ∼ r1 B and T ′ (r1 +
r2 , r3 , . . .) = B ∼ r1 A.
5.7 Exercises
Exercise 1 Show that there are 6 integral trees on at most ten vertices, namely
(i) K1 , (ii) K2 , (iii) K1,4 = D̂4 , (iv) D̂5 , (v) Ê6 , (vi) K1,9 . (For notation,
cf. §3.1.1.)
86 CHAPTER 5. TREES
s s s s
s T "s
b
s
s b Ts "

b "
s s
b"
"b "b "b
s s s
" b" b" bs
" b
s s s
" bs
" b
s s
" bs
" b
s s s
" bs
"b
s s
An integral tree on 31 vertices.
"
s bs
What is the spectrum?
s
Exercise 2 Show that the only trees that have integral Laplace spectrum are
the stars K1,m .
Chapter 6
Groups and graphs
6.1 Γ(G, H, S)
Let G be a finite group, and H a subgroup, and S a subset of G. We can
define a graph Γ(G, H, S) by taking as vertices the cosets gH (g ∈ G), and
calling g1 H and g2 H adjacent when Hg2−1 g1 H ⊆ HSH. The group G acts as
a group of automorphisms on Γ(G, H, S) via left multiplication, and this action
is transitive. The stabilizer of the vertex H is the subgroup H.
A graph Γ(G, H, S) with H = 1 is called a Cayley graph.
Conversely, let Γ be a graph with transitive group of automorphisms G. Let

x be a vertex of Γ, and let H := Gx be the stabilizer of x in G. Now Γ can be
identified with Γ(G, H, S), where S = {g ∈ G | x ∼ gx}.
If Γ is moreover edge-transitive, then S can be chosen to have cardinality 1.
Instead of representing each vertex as a coset, one can represent each vertex
y by the subgroup Gy fixing it. If H = Gx and y = gx, then Gy = gHg −1 , so
that now G acts by conjugation.
6.2 Spectrum
Let Γ be a graph and G a group of automorphisms. Let M be a matrix with
rows and columns indexed by the vertex set of Γ, and suppose that M commutes
with all elements of G (so that gM = M g, or, equivalently, Mxy = Mgx,gy ).
Now tr gM only depends on the conjugacy class of g in G, so the map g 7→ tr gM
defines a class function on G.
(Also the spectrum of gM only depends on the conjugacy class of g in G,
but it is not clear how the spectrum should be ordered. Having the trace,
however, suffices: one can retrieve the spectrum of a matrix M from the traces
i
of the powers
P sM . People also introduce the zeta function of a graph Γ by
ζΓ (−s) = λ = tr Ls , where the sum is over the eigenvalues λ of the Laplacian
L, in order to have a single object the trace of which encodes the spectrum.)
If Γ has vertex set X, and V = RX is the R-vector space spanned by the
vertices of Γ, then by Schur’s Lemma M acts as a multiple of the identity on
87
88 CHAPTER 6. GROUPS AND GRAPHS
each irreducible G-invariant subspace of V . In other words, the irreducible G-

invariant subspaces are eigenspaces of M . If M acts like θI on the irreducible
G-invariant subspace W with character χ, then tr gM |W = θχ(g).
Example Let Γ be the Petersen graph, with as vertices the unordered pairs
from a 5-set, adjacent when they are disjoint, and let M = A, the adjacency
matrix. Now f (g) := tr gA = #{x | x ∼ gx} defines a class function on
Aut Γ = Sym(5). Below we show f together with the character table of Sym(5)
(with top row indicating the cycle shape of the element):
1 2 22 3 4 5 2.3
χ1 1 1 1 1 1 1 1
χ2 1 −1 1 1 −1 1 −1
χ3 4 2 0 1 0 −1 −1
χ4 4 −2 0 1 0 −1 1
χ5 5 1 1 −1 −1 0 1
χ6 5 −1 1 −1 1 0 −1
χ7 6 0 −2 0 0 1 0
f 0 0 4 0 2 5 6
We see that f = 3χ1 − 2χ3 + χ5 . It follows that Γ has spectrum 31 (−2)4 15 ,

where the eigenvalues are the coefficients of f written as linear combination of
irreducible characters, and the multiplicities are the degrees of these characters.
The permutation character is π = χ1 + χ3 + χ5 (obtained for M = I). It is
multiplicity free, that is, no coefficients larger than 1 occur. In the general case
the coefficient of an irreducible character χ in the expression for f will be the
sum of the eigenvalues of M on the irreducible subspaces with character χ.
6.3 Covers
Let a graph Γ = (X, E) consist of a set of vertices X and a set of edges E and
an incidence relation between X and E (such that each edge is incident with
one or two points). An edge incident with one point only is called a loop. A
homomorphism f : Γ → ∆ of graphs is a map that sends vertices to vertices,
edges to edges, loops to loops, and preserves incidence.
For example, the chromatic number of Γ is the smallest integer m such that
there is a homomorphism from Γ to Km .
The map f is called a covering when it is a surjective homomorphism, and
for each vertex x of Γ and each edge e of ∆ that is incident with f (x), there is
a unique edge ẽ of Γ that is incident with x such that f (ẽ) = e. Now Γ is called
a cover of ∆.
If f is a covering, then paths in ∆ starting at a vertex y of ∆ lift uniquely
to paths starting at a vertex x of Γ, for each x ∈ f −1 (y).
The universal cover of a connected graph ∆ is the unique tree T that is a
cover. If a is a fixed vertex of ∆, then the vertices of T can be identified with
the walks in ∆ starting at a that never immediately retrace an edge, where
two walks are adjacent when one is the extension of the other by one more
edge. The tree T will be infinite when ∆ contains at least one cycle. If f is the
6.3. COVERS 89
covering map (that assigns to a walk its final vertex), then T has a group of
automorphisms H acting regularly on the fibers of f .
Given an arbitrary collection of cycles C in ∆, and a positive integer nC for
each C ∈ C, one may consider the most general cover satisfying the restriction
that the inverse image of the walk traversing C nC times is closed. (For example,
the ‘universal cover modulo triangles’ is obtained by requiring that the preimage
of each triangle is a triangle.) There is a unique such graph, quotient of the
universal cover. Again the covering group (the group preserving the fibers) acts
regularly on the fibers.
Conversely, let Γ be a graph, and H a group of automorphisms. The quotient
graph Γ/H has as vertices the H-orbits on the vertices of Γ, as edges the H-
orbits on the edges of Γ, and a vertex xH is incident with an edge eH when
some element of xH is incident with some element of eH .
The natural projection π : Γ → Γ/H is a homomorphism. It will be a
covering when no vertex x of Γ is on two edges in an orbit eH . In this case we
also say that Γ is a cover of Γ/H.
Now let Γ be finite, and f : Γ → ∆ a covering. P Let AΓ and A∆ be the
adjacency matrices of Γ and ∆. Then (A∆ )f (x),z = y∈f −1 (z) (AΓ )xy . If we
view AΓ and A∆ as linear transformations on the vector spaces VΓ and V∆
spanned by the vertices of Γ and ∆, and extend f to a linear map, then this
equation becomes A∆ ◦ f = f ◦ AΓ . If u is an eigenvector of ∆ with eigenvalue
θ, then u ◦ f (defined by (u ◦ f )y = uf (y) ) is an eigenvector of Γ with the same
eigenvalue, and the same holds for Laplacian eigenvectors and eigenvalues.
(This is immediately clear, but also follows from the fact that the partition
of V Γ into fibers f −1 (z) is an equitable partition.)
For example, let Γ be the path on 6 vertices with a loop added on both
sides and ∆ the path on 2 vertices with a loop added on both sides. Then the
map sending vertices 1, 4, 5 of Γ to one vertex of ∆ and 2, 3, 6 to the other,
is a covering. The ordinary spectrum√ of ∆ is 2,√ 0, and hence also Γ has these
eigenvalues. (It has spectrum 2, 3, 1, 0, −1, − 3.)
Thus, the spectrum of ∆ is a subset of the spectrum of Γ. We can be more
precise and indicate which subset.
Let V = RX be the vector space spanned by the vertices of Γ. Let G be
a group of automorphisms of Γ. We can view the elements g ∈ G as linear
transformations of V (permuting the basis vectors). Let H be a subgroup of G,
and let W be the subspace of V fixed by H.
Lemma 6.3.1 Let M be a linear transformation of V that commutes with all

g ∈ G. Then M preserves W and tr M |W = (1H , φM |H ) = (1G
H , φM ) where φM
is the class function on G defined by φM (g) = tr gM .
Proof. The orthogonal projection P from V onto W is given by

1 X
P = h.
|H|
h∈H
If M commutes with all h ∈ H then M P u = P M u, so M preserves the

fixed space W , and its restriction
P M |W has trace trP M . Expanding P we
1
find tr M |W = tr P M = |H| h∈H tr hM = (1H , φM |H ). The second equality
follows by Frobenius reciprocity.

Now assume that the map π : Γ → Γ/H is a covering. Then π ◦ AΓ = AΓ/H ◦ π.
One can identify the vector space VΓ/H spanned by the vertices of Γ/H with
P
the vector space W : the vertex xH corresponds to √1 h
h∈H x ∈ W . This
|H|
identification identifies AΓ/H with A|W . This means that the above lemma (ap-
plied with M = A) gives the spectrum of Γ/H. In precisely the same way, for
M = L, it gives the Laplacian spectrum of Γ/H.
We see that for a covering the spectrum of the quotient Γ/H does not depend
on the choice of H, but only on the permutation character 1G H . This is Sunada’s
observation, and has been used to construct cospectral graphs, see §13.2.4.
6.4 Cayley sum graphs

In §1.4.8 we discussed Cayley graphs for an abelian group G. A variation is
the concept of Cayley sum graph with sum set S in an abelian group G. It has
vertex set G, and two elements g, h ∈ G are adjacent when g + h ∈ S. (Other
terms are addition Cayley graphs or just sum graphs.)
It is easy to determine the spectrum of a Cayley sum graph.
Proposition 6.4.1 ([131]) Let Γ be the Cayley sum graph with sum set S in
the finite abelian group G. Let χ run through the n = |G| characters of G. The
spectrum of Γ consists of the numbers χ(S) for each real χ, P and ±|χ(S)| for
each pair χ, χ of conjugate non-real characters, where χ(S) = s∈S χ(s).
P P
Proof. If χ : G → C∗ is a character of G, then y ∼ x χ(y) = s∈S χ(s − x) =
P
( s∈S χ(s))χ(−x) = χ(S)χ(x). Now Γ is undirected, so the spectrum is real.
If χ is a real character, then we found an eigenvector χ, with eigenvalue χ(S).
If χ is non-real, then pick a constant α so that |χ(S)| = α2 χ(S). Then Re(αχ)
and Im(αχ) are eigenvectors with eigenvalues |χ(S)| and −|χ(S)|, respectively.

Chung [85] constructs Cayley sum graphs that are good expanders. For
further material on Cayley sum graphs, see [4], [83], [164], [168].
6.4.1 (3,6)-fullerenes
An amusing application was given by DeVos et al. [131]. A (3, 6)-fullerene is a
cubic plane graph whose faces (including the outer face) have sizes 3 or 6. Fowler
conjectured (cf. [146]) that such graphs have spectrum Φ ∪ {3, −1, −1, −1} (as
multiset), where Φ = −Φ, and this was proved in [131].
For example, the graph
d r rh
A A
b r Ar c f r Arg
A A
aA
r A
re
√ √
has spectrum 3, 5, 1,√(−1)4 , − 5 with eigenvalues 3, −1, −1, −1 together with
the symmetric part ± 5, ±1.
6.5. EXERCISES 91
The proof goes as follows. Construct the bipartite double Γ ⊗ K2 of Γ. This

is a cover of Γ, and both triangles and hexagons lift to hexagons, three at each
vertex, so that Γ ⊗ K2 is a quotient of H, the regular tesselation of the plane
with hexagons.
fq sg
A
aq se eAq sa
A A
cq sb gAq sf cAq sb
A A A
hq sd dAq sh hAq sd dAq sh
A A A A
gs bAq sc fAq sg bAq sc Aqf
A A A A
eAq
sa aAq se eAq sa aAq e
s
A A A
cAq sb gAq sf cAq sb
A A
dAq sh hAq sd
A
fAq sg
Let H have set of vertices H, and let Γ ⊗ K2 have vertex set U , and let Γ
have vertex set V . Let π : H → U and ρ : U → V be the quotient maps. The
graph Γ ⊗ K2 is bipartite with bipartite halves U1 and U2 , say. Fix a vertex
a1 ∈ U1 and call it 0. Now π −1 (U1 ) is a lattice in R2 , and π −1 (a1 ) is a sublattice
(because the concatenation of two walks of even length in Γ starting and ending
in a again is such a walk), so the quotient G = π −1 (U1 )/π −1 (a1 ) is an abelian
group, and G can be naturally identified with V . The automorphism of Γ ⊗ K2
that for each u ∈ V interchanges the two vertices u1 , u2 of ρ−1 (u), lifts (for each
choice of a ∈ π −1 (a2 )) to an isometry of H with itself that is a point reflection
x 7→ v − x (where v = a). It follows that if two edges x1 y2 and z1 w2 in H are
parallel, then x + y = z + w. Hence Γ is the Cayley sum graph for G where the
sum set S is the set of three neighbors of a in Γ.
Now the spectrum follows. By the foregoing, the spectrum consists of the
values ±|χ(S)| for non-real characters χ of G, and χ(S) for real characters.
Since tr A = 0 and Γ is cubic and not bipartite (it has four triangles) it suffices
to show that there are precisely four real characters (then the corresponding
eigenvalues must be 3, −1, −1, −1). But this is clear since the number of real
characters is the number of elements of order 2 in G, an abelian group with
(at most) two generators, hence at most four, and fewer than four would force
nonzero tr A. This proves Fowler’s conjecture.
6.5 Exercises
Exercise 1 Show that a (3,6)-fullerene has precisely four triangles.
Chapter 7
Euclidean representations
The main goal of this chapter is the famous result by Cameron, Goethals, Seidel
and Shult [77] characterizing graphs with smallest eigenvalue not less than −2.
7.1 Examples
We have seen examples of graphs with smallest eigenvalue θmin ≥ −2. The most
important example is formed by the line graphs (see §1.4.5), and people wanted
to characterize line graphs by this condition and possibly some additional hy-
potheses.
Another series of examples are the so-called cocktailparty graphs, that is, the
graphs Km×2 , i.e., mK2 , with spectrum 2m − 2, 0m , (−2)m−1 . For m ≥ 4 these
are not line graphs.
And there are exeptional examples like the Petersen graph (with spectrum
3 15 (−2)4 ), lots of them. It is easy to see that the Petersen graph is not a line
graph. More generally, no line graph can have a 3-claw, that is, an induced K1,3
subgraph, as is immediately clear from the definition.
7.2 Euclidean representation

Now suppose Γ has smallest eigenvalue θmin ≥ −2. Then A + 2I is positive
semidefinite, so that A + 2I is the Gram matrix of a collection of vectors in
some Euclidean space Rm (where m = rk (A + 2I)), cf. §2.9.
In this way we obtain a map x 7→ x̄ from vertices of Γ to vectors in Rm ,
where 
 2 if x = y
(x̄, ȳ) = 1 if x ∼ y

0 if x 6∼ y.
The additive subgroup of Rm generated by the vectors x̄, for x in the vertex
set X of Γ, is a root lattice: an integral lattice generated by roots: vectors with
squared length 2. Root lattices have been classified. That classification is the
subject of the next section.
93
94 CHAPTER 7. EUCLIDEAN REPRESENTATIONS
7.3 Root lattices

We start with an extremely short introduction into lattices.
Lattice
A lattice Λ is a discrete additive subgroup of Rn . Equivalently, it is a finitely
generated free Z-module with positive definite symmetric bilinear form.
Basis
Assume that our lattice Λ has dimension n, i.e., spans Rn . Let {a1 , ..., an } be
a Z-basis of Λ. Let A be the matrix with the P vectors ai as rows. If we choose
a different Z-basis {b1 , ..., bn }, so that bi = sij aj , and B is the matrix with
the vectors bi as rows, then B = SA, with S = (sij ). Since S is integral and
invertible, it has determinant ±1. It follows that | det A| is uniquely determined
by Λ, independent of the choice of basis.
Volume
Rn /Λ is an n-dimensional torus, compact with finite volume. Its volume is the
volume of the fundamental domain, which equals | det A|.
If Λ′ is a sublattice of Λ, then vol(Rn /Λ′ ) = vol(Rn /Λ).|Λ/Λ′ |.
Gram matrix
Let G be the matrix (ai , aj ) of inner
√ products of basis vectors for a given basis.
Then G = AA⊤ , so vol(Rn /Λ) = det G.
Dual lattice
The dual Λ∗ of a lattice Λ is the lattice of vectors having integral inner products
with all vectors in Λ: Λ∗ = {x ∈ Rn | (x, r) ∈ Z for all r ∈ Λ}.
It has a basis {a∗1 , ..., a∗n } defined by (a∗i , aj ) = δij .
Now A∗ A⊤ = I, so A∗ = (A−1 )⊤ and Λ∗ has Gram matrix G∗ = G−1 .
It follows that vol(Rn /Λ∗ ) = vol(Rn /Λ)−1 .
We have Λ∗∗ = Λ.
Integral lattice
The lattice Λ is called integral when every two lattice vectors have an integral
inner product.
For an integral lattice Λ one has Λ ⊆ Λ∗ .
The lattice Λ is called even when (x, x) is an even integer for each x ∈ Λ.
An even lattice is integral.
7.3. ROOT LATTICES 95
Discriminant
The determinant, or discriminant, disc Λ of a lattice Λ is defined by disc Λ =
det G. When Λ is integral, we have disc Λ = |Λ∗ /Λ|.
A lattice is called self-dual or unimodular when Λ = Λ∗ , i.e., when it is
integral with discriminant 1. An even unimodular lattice is called Type II, the
remaining unimodular lattices are called Type I.
It can be shown that if there is an even unimodular lattice in Rn , then n is
divisible by 8.
Direct sums
If Λ and Λ′ are lattices in Rm and Rn , respectively, then Λ ⊥ Λ′ , the orthogonal
direct sum of Λ and Λ′ , is the lattice {(x, y) ∈ Rm+n | x ∈ Λ and y ∈ Λ′ }.
A lattice is called irreducible when it is not the orthogonal direct sum of two
nonzero lattices.
Examples
Zn
r r r r
r r r r
r r r r
The lattice Zn is unimodular, type I.
A2
r r r
s ` 2r + s
q
`
p
r r r
0 r
r r r
The triangular lattice in the plane R2 has basis {r,s}. Choose the scale such
√ 2 −1

that r has length 2. Then the Gram matrix is G = −1 2 , so that det G = 3
and p, q ∈ A∗2 . A fundamental region for A2 is the parallelogram on 0, r, s. A
fundamental region for A∗2 is the parallelogram on 0, p, q. Note that the area of
the former is thrice that of the latter.
2
The representation of this lattice in RP has nonintegral coordinates. It is
3
easier to work in R , on the hyperplane xi = 0, and choose r = (1, −1, 0),
sP= (0, 1, −1). Then A2 consists of the points (x1 , x2 , x3 ) with xi ∈ Z and
∗
xi = 0. The dual lattice P A2 consists of the points (x1 , x2 , x3 ) with x1 ≡
x2 ≡ x3 ≡ 0 (mod 1) and xi = 0 (so that 3x1 ∈ Z). It contains for example
p = 13 (2r + s) = ( 23 , − 13 , − 13 ).
E8
Let ρ : Zn → 2n be coordinatewise reduction mod 2. Given a binary linear
code C, the lattice ρ−1 (C) is integral, since it is contained in Zn , but never
unimodular, unless it is all of Zn , a boring situation.
Now suppose that C is self-orthogonal, so that any two code words have an
even inner product. Then √12 ρ−1 (C) is an integral lattice. If dim C = k then
1
vol(Rn /ρ−1 (C)) = 2n−k and hence vol(Rn / √12 ρ−1 (C)) = 2 2 n−k . In particular,
√1 ρ−1 (C) will be unimodular when C is self-dual, and even when C is ‘doubly
2
even’, i.e., has weights divisible by 4.
Let C be the [8,4,4] extended Hamming code. Then √12 ρ−1 (C) is an even
unimodular 8-dimensional lattice known as E8 .
The code C has weight enumerator 1 + 14X 4 + X 8 (that is, has one word
of weight 0, 14 words of weight 4, and one word of weight 8). It follows that
the roots (vectors r with (r, r) = 2) in this incarnation of E8 are the 16 vectors
± √12 (2, 0, 0, 0, 0, 0, 0, 0) (with 2 in any position), and the 16.14 = 224 vectors
√1 (±1, ±1, ±1, ±1, 0, 0, 0, 0) with ±1 in the nonzero positions of a weight 4
2
vector. Thus, there are 240 roots.
Root lattices
A root lattice is an integral lattice generated by roots (vectors r with (r, r) = 2).
For example, A2 and E8 are root lattices.
The set of roots in a root lattice is a (reduced) root system Φ, i.e., satisfies
(i) If r ∈ Φ and λr ∈ Φ, then λ = ±1.
(r,s)
(ii) Φ is closed under the reflection wr that sends s to s − 2 (r,r) r for each
r ∈ Φ.
(r,s)
(iii) 2 (r,r) ∈ Z.
Since Φ generates Λ and Φ is invariant under W = hwr | r ∈ Φi, the same
holds for Λ, so root lattices have a large group of automorphisms.
A fundamental system of roots Π in a root lattice Λ is a set of roots generating
Λ and such that (r, s) ≤ 0 for distinct r, s ∈ Π. A reduced fundamental system
of roots is a fundamental system that is linearly independent. A non-reduced
fundamental system is called extended.
For example, in A2 the set {r, s} is a reduced fundamental system, and
{r, s, −r − s} is an extended fundamental system.
The Dynkin diagram of a fundamental system Π such that (r, s) 6= −2 for
r, s ∈ Π, is the graph with vertex set Π where r and s are joined by an edge
when (r, s) = −1. (The case (r, s) = −2 happens only for a non-reduced system
with A1 component. In that case we do not define the Dynkin diagram.)
Every root lattice has a reduced fundamental system: Fix some vector u,
not orthogonal to any root. Put Φ+ (u) = {r ∈ Φ | (r, u) > 0} and Π(u) =
{r ∈ Φ+ (u) | r cannot be written as s + t with s, t ∈ Φ+ (u)}. Then Π(u) is a
reduced fundamental system of roots, and written on this basis each root has
only positive or only negative coefficients.
(Indeed, if r, s ∈ Π(u) and (r, s) = 1, then say r − s ∈ Φ+ (u) and r =
(r − s) + s, contradiction. This shows that Π(u) is a fundamental system. If
7.3. ROOT LATTICES 97
P P
P γr r = 0, then separate the γr into positive and negative ones to get αr r =
P sβ s = x 6
= 0 where all coefficients α r , β s are positive. Now 0 < (x, x) =
αr βs (r, s) < 0, contradiction. This shows that Π(u) is reduced. Now also the
last claim follows.)
Proposition 7.3.1 Let Π be a reduced fundamental system.

(i) For all x ∈ Rn there is a w ∈ W such that (w(x), r) ≥ 0 for all r ∈ Π.
(ii) Π = Π(u) for some u. (That is, W is transitive on reduced fundamental
systems.)
(iii) If Λ is irreducible, then there is a unique r̃ ∈ Φ such that Π ∪ {r̃} is an
extended fundamental system.
Proof. (i) Let G be the Gram matrix of Π, and write A = 2I − G. Since G is

positive definite, A has largest eigenvalue less than 2. Using Perron-Frobenius,
let γ = (γr )r∈Π be a positive eigenvector of A. If (x, s) < 0 for some s ∈ Π,
then put x′ = ws (x) = x − (x, s)s. Now
X X X
(x′ , γr r) = (x, γr r) − (Gγ)s (x, s) > (x, γr r).
r r r
But W is finite, so after finitely many steps we reach the desired conclusion.
(ii) Induction on |Π|. Fix x with (x, r) ≥ 0 for all r ∈ Π. Then Π0 = Π ∩ x⊥
is a fundamental system of a lattice in a lower-dimensional space, so of the form
Π0 = Π0 (u0 ). Take u = x + εu0 for small ε > 0. Then Π = Π(u).
(iii) If r ∈ Φ+ (u) has maximal (r, u), then r̃ = −r is the unique root that can
be added. It can be added, since (r̃, s) ≥ 0 means (r, s) < 0, so that r+s is a root,
contradicting maximality of r. And it is unique because linear dependencies
of an extended system correspond to an eigenvector with eigenvalue 2 of the
extended Dynkin diagram, and by Perron-Frobenius up to a constant there is
a unique such eigenvector when the diagram is connected, that is, when Λ is
irreducible.
Classification
The irreducible root lattices one finds are An (n ≥ 0), Dn (n ≥ 4), E6 , E7 , E8 .
Each is defined by its Dynkin diagram.
P
(1) An : The lattice vectors are: x ∈ Zn+1 with xi = 0. There are n(n + 1)
roots: ei − ej (i 6= j). The discriminant is n + 1, and A∗n /An ∼
= Zn+1 , with the
1
quotient generated by n+1 (e1 + ... + en − nen+1 ) ∈ A∗n .
(e1 − en+1 )
HH
c
HH
HH
s
s s s H Hs
e2 − e1 e3 − e2 en+1 − en
P
(2) Dn : The lattice vectors are: x ∈ Zn with xi ≡ 0 (mod 2). There
are 2n(n − 1) roots ±ei ± ej (i 6= j). The discriminant is 4, and Dn∗ /Dn is
isomorphic to Z4 when n is odd, and to Z2 × Z2 when n is even. Dn∗ contains
e1 and 21 (e1 + ... + en ). Note that D3 ∼
= A3 .
e1 + e2 (−en−1 − en )
s c
s s s s s s
e2 − e1 e3 − e2 en−1 − en−2 en − en−1
(3) E8 : (Recall that we already gave a construction of E8 from the Hamming

code.) The lattice is the span of D8 and c := 12 (e1 + ... + e8 ). There are
240 = 112 + 128 roots, of the forms ±ei ± ej (i 6= j) and 21 (±e1 ± ... ± e8 ) with
an even number of minus signs. The discriminant is 1, and E8∗ = E8 .
1
2
(1, 1, 1, −1, −1, −1, −1, 1)
s
s s s s s s s c
e2 − e1 e3 − e2 e4 − e3 e5 − e4 e6 − e5 e7 − e6 −e7 − e8 (c)
(4) E7 : Take E7 = E8 ∩c⊥ . There are 126 = 56+70 roots. The discriminant
is 2, and E7∗ contains 41 (1, 1, 1, 1, 1, 1, −3, −3).
1
2
(1, 1, 1, −1, −1, −1, −1, 1)
s
c s s s s s s
(e1 − e8 ) e2 − e1 e3 − e2 e4 − e3 e5 − e4 e6 − e5 e7 − e6
(5) E6 : For the vector d = −e7 − e8 , take E6 = E8 ∩ {c, d}⊥ . There

are 72 = 32 + 40 roots. The discriminant is 3, and E6∗ contains the vector
1
3 (1, 1, 1, 1, −2, −2, 0, 0).
c (e7 − e8 )
1
s 2
(1, 1, 1, −1, −1, −1, −1, 1)
s s s s s
e2 − e1 e3 − e2 e4 − e3 e5 − e4 e6 − e5
That this is all, is an easy consequence of the Perron-Frobenius theorem: A =

2I − G is the adjacency matrix of a graph, namely the Dynkin diagram, and
this graph has largest eigenvalue at most 2. These graphs were determined in
Theorem 3.1.3. The connected graphs with largest eigenvalue less than 2 are the
Dynkin diagrams of reduced fundamental systems of irreducible root systems
and the connected graphs with largest eigenvalue 2 are the Dynkin diagrams of
extended root systems.
In the pictures above, the reduced fundamental systems were drawn with
black dots, and the additional element of the extended system with an open dot
(and a name given in parentheses).
7.4. CAMERON-GOETHALS-SEIDEL-SHULT 99
7.4 Cameron-Goethals-Seidel-Shult
Now return to the discussion of connected graphs Γ with smallest eigenvalue
θmin ≥ −2. In §7.2 we found a map x 7→ x̄ from the vertex set X of Γ to some
Euclidean space Rm such that the inner product (x̄, ȳ) takes the value 2, 1, 0
when x = y, x ∼ y and x 6∼ y, respectively.
Let Σ be the image of X under this map. Then Σ generates a root lattice
Λ. Since Γ is connected, the root lattice is irreducible.
By the classification of root lattices, it follows that the root lattice is one
of An , Dn , E6 , E7 or E8 . Note that the graph is determined by Σ, so that
the classification of graphs with θmin ≥ −2 is equivalent to the classification of
subsets Φ of the root system with the property that all inner products are 2, 1,
or 0, i.e., nonnegative.
Now An and Dn can be chosen to have integral coordinates, and E6 ⊂ E7 ⊂
E8 , so we have the two cases (i) Σ ⊂ Zm+1 , and (ii) Σ ⊂ E8 . A graph is called
exceptional in case (ii). Since E8 has a finite number of roots, there are only
finitely many exceptional graphs.
In case (i) one quickly sees what the structure of Γ has to be. Something
like a line graph with attached cocktailparty graphs. This structure has been
baptised generalized line graph. The precise definition will be clear from the
proof of the theorem below.
Theorem 7.4.1 (i) Let Γ be a connected graph with smallest eigenvalue θmin ≥
−2. Then either Γ is a generalized line graph, or Γ is one of finitely many
exceptions, represented by roots in the E8 lattice.
(ii) A regular generalized line graph is either a line graph or a cocktailparty
graph.
(iii) A graph represented by roots in the E8 lattice has at most 36 vertices,
and every vertex has valency at most 28.
Proof. (i) Consider the case Σ ⊂ Zm+1 . Roots in Zm+1 have shape ±ei ± ej .
If some ei has the same sign in all σ ∈ Σ in which it occurs, then choose
the basis such that this sign is +. Let I be the set of all such indices i. Then
{x | x̄ = ei +ej for some i, j ∈ I} induced a line graph in Γ, with x corresponding
to the edge ij on I. If j ∈ / I, then ej occurs with both signs, and there are
σ, τ ∈ Σ with σ = ±ei + ej and τ = ±ei′ − ej . Since all inner products in Σ
are nonnegative, i = i′ with i ∈ I, and σ = ei + ej , τ = ei − ej . Thus, i is
determined by j and we have a map φ : j 7→ i from indices outside I to indices
in I. Now for each i ∈ I the set {x | x̄ = ei ± ej for some j with φ(j) = i}
induces a cocktailparty graph. Altogether we see in what way Γ is a line graph
with attached cocktailparty graphs.
(ii) Now let Γ be regular. A vertex x with x̄ = ei −ej is adjacent to all vertices
with image ei ± ek different from ei + ej . But a vertex y with ȳ = ei + ek where
i, k ∈ I is adjacent to all vertices with image ei ± ek without exception (and also
to vertices with image ek ± el ). Since Γ is regular both types of vertices cannot
occur together, so that Γ is either a line graph or a cocktailparty graph.
(iii) Suppose Σ ⊂ E8 . Consider the 36-dimensional space of symmetric 8 × 8
matrices, equipped with the positive definite inner product (P, Q) = tr P Q.
Associated with the 240 roots r of E8 are 120 rank 1 matrices Pr = rr⊤ with
mutual inner products (Pr , Ps ) = tr rr⊤ ss⊤ = (r, s)2 . The Gram matrix of the
set of Pr for r ∈ Σ is G = 4I + A. Since G is positive definite (it has smallest

eigenvalue ≥ 2), the vectors Pr are linearly independent, and hence |Σ| ≤ 36.
Finally, let r be a root of E8 . The 56 roots s of E8 that satisfy (r, s) = 1 fall
into 28 pairs s, s′ where (s, s′ ) = −1. So, Σ can contain at most one member
from each of these pairs, and each vertex of Γ has valency at most 28.
The bounds in (ii) are best possible: Take the graph K8 + L(K8 ) and add edges
joining i ∈ K8 with jk ∈ L(K8 ) whenever i, j, k are distinct. This graph has 36
vertices, the vertices in K8 have 28 neighbours, and the smallest eigenvalue is
−2. A representation in E8 is given by i 7→ 12 (e1 + ... + e8 ) − ei and jk 7→ ej + ek .
There is a large amount of literature on exceptional graphs.
7.5 Exercises
Exercise 1 Show that the following describes a root system of type E6 . Take
the following 72 vectors in R9 : 18 vectors ±(u, 0, 0), ±(0, u, 0), ±(0, 0, u) with
u ∈ {(1, −1, 0), (0, 1, −1), (−1, 0, 1)}, and 54 vectors ±(u, v, w) with u, v, w ∈
{( 23 , − 13 , − 13 ), (− 13 , 32 , − 13 ), (− 13 , − 13 , 32 )}.
Exercise 2 Show that the following describes a root system of type E7 . Take
the following 126 vectors in R7 : 60 vectors ±ei ± ej with 1 ≤ i < j ≤ 6, and 64
vectors ±(x1 , ..., x6 , √12 ) with xi = ± 21 where an even number of xi has + sign,
√
and 2 vectors ±(0, ..., 0, 2).
Chapter 8
Strongly regular graphs
8.1 Strongly regular graphs

A graph (simple, undirected and loopless) of order v is called strongly regular
with parameters v, k, λ, µ whenever it is not complete or edgeless and
(i) each vertex is adjacent to k vertices,
(ii) for each pair of adjacent vertices there are λ vertices adjacent to both,
(iii) for each pair of non-adjacent vertices there are µ vertices adjacent to both.
We require that both edges and non-edges occur, so that the parameters are
well-defined.
In association scheme terminology (cf. §10.1), a strongly regular graph is
a symmetric association scheme with two (nonidentity) classes, in which one
relation is singled out to be the adjacency relation.
8.1.1 Simple examples

Easy examples of strongly regular graphs:
(i) A quadrangle is strongly regular with parameters (4, 2, 0, 2).
(ii) A pentagon is strongly regular with parameters (5, 2, 0, 1).
(iii) The 3×3 grid, the Cartesian product of two triangles, is strongly regular
with parameters (9, 4, 1, 2).
(iv) The Petersen graph is strongly regular with parameters (10, 3, 0, 1).
(Each of these graphs is uniquely determined by its parameters, so if you do
not know what a pentagon is, or what the Petersen graph is, this defines it.)
Each of these examples can be generalized in numerous ways. For example,
(v) Let q = 4t + 1 be a prime power. The Paley graph Paley(q) is the
graph with the finite field Fq as vertex set, where two vertices are adjacent
when they differ by a (nonzero) square. It is strongly regular with parameters
(4t + 1, 2t, t − 1, t), as we shall see below. Doing this for q = 5 and q = 9, we
find the examples (ii) and (iii) again. For q = 13 we find a graph that is locally
a hexagon. For q = 17 we find a graph that is locally an 8-gon + diagonals.
101
102 CHAPTER 8. STRONGLY REGULAR GRAPHS
(vi) The m × m grid, the Cartesian product of two complete graphs on m

vertices, is strongly regular with parameters (m2 , 2(m−1), m−2, 2) (for m > 1).
For m = 2 and m = 3 we find the examples (i) and (iii) again.
(vii) The complete multipartite graph Km×a , with vertex set partitioned into
m groups of size a, where two points are adjacent when they are from different
groups, is strongly regular with parameters (ma, (m − 1)a, (m − 2)a, (m − 1)a)
(for m > 1 and a > 1). For m = a = 2 we find Example (i) again.
The complement of a graph Γ is the graph Γ̄ with the same vertex set as Γ,
where two vertices are adjacent if and only if they are nonadjacent in Γ. The
complement of a strongly regular graph with parameters (v, k, λ, µ) is again
strongly regular, and has parameters (v, v − k − 1, v − 2k + µ − 2, v − 2k + λ).
(Indeed, we keep the same association scheme, but now single out the other
nonidentity relation.)
(viii) The Paley graph Paley(q) is isomorphic to its complement. (Indeed,
an isomorphism is given by multiplication by a nonsquare.) In particular we see
that the pentagon and the 3×3 grid are (isomorphic to) their own complements.
(ix) The disjoint union mKa of m complete graphs of size a is strongly
regular with parameters (ma, a − 1, a − 2, 0) (for m > 1 and a > 1). These
graphs are the complements of those in Example (vii).
(x) The triangular graph on the pairs in an m-set, denoted by T (m), or
by m 2 , has these pairs as vertices, where two pairs are adjacent whenever
they meet in one point. These graphs are strongly regular, with parameters
( m
2 , 2(m − 2), m − 2, 4), if m ≥ 4. For m = 4 we find K3×2 . For m = 5 we find
the complement of the Petersen graph.
The four parameters are not independent. Indeed, if µ 6= 0 we find the
relation
k(k − 1 − λ)
v =1+k+
µ
by counting vertices at distance 0, 1 and 2 from a given vertex.
8.1.2 The Paley graphs

Above we claimed that the Paley graphs (with vertex set Fq , where q is a
prime power congruent 1 mod 4, and where two vertices are adjacent when
their difference is a nonzero square) are strongly regular. Let us verify this.
Proposition 8.1.1 The Paley graph Paley(q) with q = 4t+1 is strongly regular
with parameters (v, k, λ, µ) = (4t + 1, 2t, t − 1, t). It has eigenvalues k, (−1 ±
√
q)/2 with multiplicities 1, 2t, 2t, respectively.
Proof. The values for v and k are clear. Let χ : Fq → {−1, 0, 1} be the
quadratic residue character defined by χ(0) =P0, χ(x) = 1 when x is a (nonzero)
square, andPχ(x) = −1 otherwise.
P Note that x χ(x) = 0, and that for nonzero
a we have z χ(z 2 − az) = z6=0 χ(1 − az ) = −1. Now λ and µ follow from
X X
4 1= (χ(z − x) + 1)(χ(z − y) + 1) = −1 − 2χ(x − y) + (q − 2).
z
x∼z∼y
z6=x,y
For the spectrum, see Theorem 8.1.3 below.

8.1. STRONGLY REGULAR GRAPHS 103
8.1.3 Adjacency matrix

For convenience we call an eigenvalue restricted if it has an eigenvector perpen-
dicular to the all-ones vector 1.
Theorem 8.1.2 For a simple graph Γ of order v, not complete or edgeless, with
adjacency matrix A, the following are equivalent:
(i) Γ is strongly regular with parameters (v, k, λ, µ) for certain integers k, λ,
µ,
(ii) A2 = (λ − µ)A + (k − µ)I + µJ for certain real numbers k, λ, µ,
(iii) A has precisely two distinct restricted eigenvalues.
Proof. The equation in (ii) can be rewritten as
A2 = kI + λA + µ(J − I − A).
Now (i) ⇐⇒ (ii) is obvious.

(ii) ⇒ (iii): Let ρ be a restricted eigenvalue, and u a corresponding eigenvector
perpendicular to 1. Then Ju = 0. Multiplying the equation in (ii) on the right
by u yields ρ2 = (λ−µ)ρ+(k −µ). This quadratic equation in ρ has two distinct
solutions. (Indeed, (λ − µ)2 = 4(µ − k) is impossible since µ ≤ k and λ ≤ k − 1.)
(iii) ⇒ (ii): Let r and s be the restricted eigenvalues. Then (A−rI)(A−sI) = αJ
for some real number α. So A2 is a linear combination of A, I and J.
8.1.4 Imprimitive graphs

A strongly regular graph is called imprimitive if it, or its complement, is dis-
connected, and primitive otherwise. Imprimitive strongly regular graphs are
boring.
If a strongly regular graph is not connected, then µ = 0 and k = λ + 1. And
conversely, if µ = 0 or k = λ + 1 then the graph is a disjoint union aKm of some
number a of complete graphs Km . In this case v = am, k = m − 1, λ = m − 2,
µ = 0 and the spectrum is (m − 1)a , (−1)a(m−1) .
If the complement of a strongly regular graph is not connected, then k = µ.
And conversely, if k = µ then the graph is the complete multipartite graph
Ka×m , the complement of aKm , with parameters v = am, k = µ = (a − 1)m,
λ = (a − 2)m and spectrum (a − 1)m1 , 0a(m−1) , (−m)a−1 .
Let r and s (r > s) be the restricted eigenvalues of A. For a primitive
strongly regular graph one has k > r > 0 and s < −1.
8.1.5 Parameters
Theorem 8.1.3 Let Γ be a strongly regular graph with adjacency matrix A and
parameters (v, k, λ, µ). Let r and s (r > s) be the restricted eigenvalues of A
and let f , g be their respective multiplicities. Then
(i) k(k − 1 − λ) = µ(v − k − 1),
(ii) rs = µ − k, r + s = λ − µ,
(r+s)(v−1)+2k
(iii) f, g = 21 (v − 1 ∓ r−s ).
(iv) If r and s are non-integral, then f = g and (v, k, λ, µ) = (4t + 1, 2t, t − 1, t)

for some integer t.
Proof. (i) Fix a vertex x of Γ. Let Γ(x) and ∆(x) be the sets of vertices
adjacent and non-adjacent to x, respectively. Counting in two ways the number
of edges between Γ(x) and ∆(x) yields (i). The equations (ii) are direct conse-
quences of Theorem 8.1.2(ii), as we saw in the proof. Formula (iii) follows from
f +g = v −1 and 0 = trace A = k +f r+gs = k + 21 (r+s)(f +g)+ 21 (r−s)(f −g).
Finally, when f 6= g then one can solve for r and s in (iii) (using (ii)) and find
that r and s are rational, and hence integral. But f = g implies (µ − λ)(v − 1) =
2k, which is possible only for µ − λ = 1, v = 2k + 1.
These relations imply restrictions for the possible values of the parameters.
Clearly, the right hand sides of (iii) must be positive integers. These are the
so-called rationality conditions.
8.1.6 The half case and cyclic strongly regular graphs

The case of a strongly regular graph with parameters (v, k, λ, µ) = (4t+1, 2t, t−
1, t) for some integer t is called the half case. Such graphs are also called
conference graphs. If such a graph exists, then v is the sum of two squares, see
Theorem 9.4.2 below. The Paley graphs (§8.1.2, §9.4, §12.6) belong to this case,
but there are many further examples.
A characterization of the Paley graphs of prime order is given by
Proposition 8.1.4 (Kelly [214], Bridges & Mena [38]) A strongly regular graph
with a regular cyclic group of automorphisms is a Paley graph with a prime
number of vertices.
(See the discussion of translation association schemes in BCN [48], §2.10.

This result has been rediscovered several times.)
8.1.7 Strongly regular graphs without triangles

As an example of the application of the rationality conditions we classify the
strongly regular graphs of girth 5.
Theorem 8.1.5 (Hoffman & Singleton [196]) Suppose (v, k, 0, 1) is the param-
eter set of a strongly regular graph. Then (v, k) = (5, 2), (10, 3), (50, 7) or
(3250, 57).
Proof. The rationality conditions imply that either f = g, which leads to

(v, k) = (5, 2), or r − s is an integer dividing (r + s)(v − 1) + 2k. By Theorem
8.1.3(i)–(ii) we have
s = −r − 1, k = r2 + r + 1, v = r4 + 2r3 + 3r2 + 2r + 2,
and thus we obtain r = 1, 2 or 7.

The first three possibilities are uniquely realized by the pentagon, the Petersen
graph and the Hoffman-Singleton graph. For the last case existence is unknown
(but see §10.5.1).
More generally we can look at strongly regular graphs of girth at least 4. Seven
examples are known.
(i) The pentagon, with parameters (5,2,0,1).
(ii) The Petersen graph, with parameters (10,3,0,1). This is the complement
of the triangular graph T (5).
(iii) The folded 5-cube, with parameters (16,5,0,2). This graph is obtained
from the 5-cube 25 on 32 vertices by identifying antipodal vertices. (The com-
plement of this graph is known as the Clebsch graph.)
(iv) The Hoffman-Singleton graph, with parameters (50,7,0,1). There are
many constructions for this graph, cf., e.g., [48], §13.1. A short one, due to N.
Robertson, is the following. Take 25 vertices (i, j) and 25 vertices (i, j)′ with
i, j ∈ Z5 , and join (i, j) with (i, j + 1), (i, j)′ with (i, j + 2)′ , and (i, k) with
(j, ij + k)′ for all i, j, k ∈ Z5 . Now the subsets (i, ∗) become pentagons, the
(i, ∗)′ become pentagons (drawn as pentagrams), and each of the 25 unions of
(i, ∗) with (j, ∗)′ induces a Petersen subgraph.
(v) The Gewirtz graph, with parameters (56,10,0,2). This is the graph with
as vertices the 77 − 21 = 56 blocks of the unique Steiner system S(3, 6, 22)
not containing a given symbol, where two blocks are adjacent when they are
disjoint. It is a subgraph of the following.
(vi) The M22 graph, with parameters (77,16,0,4). This is the graph with as
vertices the 77 blocks of the unique Steiner system S(3, 6, 22), adjacent when
they are disjoint. It is a subgraph of the following.
(vii) The Higman-Sims graph, with parameters (100,22,0,6). This is the
graph with as 1 + 22 + 77 vertices an element ∞, the 22 symbols of S(3, 6, 22),
and the 77 blocks of S(3, 6, 22). The element ∞ is adjacent to the 22 symbols,
each symbol is adjacent to the 21 blocks containing it, and blocks are adjacent
when disjoint. The (rank 3) automorphism group of this graph is HS.2, where
HS is the sporadic simple group of Higman and Sims. This graph can be
partitioned into two halves, each inducing a Hoffman-Singleton graph, cf. [48],
§13.1.
Each of these seven graphs is uniquely determined by its parameters. It is
unknown whether there are any further examples. There are infinitely many
feasible parameter sets. For the parameters (324, 57, 0, 12) nonexistence was
shown in Gavrilyuk & Makhnev [150] and in Kaski & Östergård [213].
8.1.8 Further parameter restrictions

Except for the rationality conditions, a few other restrictions on the parameters
are known. We mention two of them. The Krein conditions, due to Scott [270],
can be stated as follows:
(r + 1)(k + r + 2rs) ≤ (k + r)(s + 1)2 ,
(s + 1)(k + s + 2rs) ≤ (k + s)(r + 1)2 .

When equality holds in one of these, the subconstituents of the graph (the
induced subgraphs on the neighbours and on the nonneighbours of a given point)
are both strongly regular (in the wide sense) again. For example, in the Higman-
Sims graph with parameters (v, k, λ, µ) = (100, 22, 0, 6) and k, r, s = 22, 2, −8
the second subconstituent of any point has parameters (77, 16, 0, 4).
Seidel’s absolute bound for the number of vertices of a primitive strongly regular
graph (see Corollary 9.6.8 below) reads
v ≤ f (f + 3)/2, v ≤ g(g + 3)/2.
For example, the parameter set (28,9,0,4) (spectrum 91 121 (−5)6 ) is ruled
out both by the second Krein condition and by the absolute bound.
A useful identity is an expression for the Frame quotient (cf. [48], 2.2.4 and
2.7.2). One has
f g(r − s)2 = vk(v − 1 − k).
(as is easy to check directly from the expressions for f and g given in Theorem
8.1.3(iii)). From this one immediately concludes that if v is prime, then r − s =
√
v and we are in the ‘half case’ (v, k, λ, µ) = (4t + 1, 2t, t − 1, t).
The Frame quotient, Krein conditions and absolute bound are special cases of
general (in)equalities for association schemes—see also §10.4 below. In Brouwer
& van Lint [56] one may find a list of known restrictions and constructions. It
is a sequel to Hubaut’s [202] earlier survey of constructions.
Using the above parameter conditions, Neumaier [249] derives the µ-bound:
Theorem 8.1.6 For a primitive strongly regular graph µ ≤ s3 (2s+3). If equal-
ity holds, then r = −s2 (2s + 3).
Examples of equality in the µ-bound are known for s = −2 (the Schläfli
graph, with (v, k, λ, µ) = (27, 16, 10, 8)) and s = −3 (the McLaughlin graph,
with (v, k, λ, µ) = (275, 162, 105, 81)).
Brouwer & Neumaier [58] showed that a connected partial linear space with
girth at least 5 and more than one line, in which every point is collinear with
m other points, contains at least 21 m(m + 3) points. It follows that a strongly
regular graph with µ = 2 either has k ≥ 12 λ(λ + 3) or has (λ + 1)|k.
Bagchi [12] showed that any K1,1,2 -free strongly regular graph is either the
collinearity graph of a generalized quadrangle or satisfies k ≥ (λ + 1)(λ + 2).
(It follows that in the above condition on µ = 2 the (λ + 1)|k alternative only
occurs for the m × m grid, where m = λ + 2.)
8.1.9 Strongly regular graphs from permutation groups

Suppose G is a permutation group, acting on a set Ω. The rank of the action
is the number of orbits of G on Ω × Ω. (These latter orbits are called orbitals.)
If R is an orbital, or a union of orbitals, then (Ω, R) is a directed graph that
admits G as group of automorphisms.
If G is transitive of rank 3 and its orbitals are symmetric (for all x, y ∈ Ω
the pairs (x, y) and (y, x) belong to the same orbital), say with orbitals I, R, S,
where I = {(x, x) | x ∈ Ω}, then (Ω, R) and (Ω, S) is a pair of complementary
strongly regular graphs.
For example, let G be Sym(n) acting on a set Σ of size 5. This action induces
an action on the set Ω of unordered pairs of elements in Σ, and this latter action
is rank 3, and gives the pair of graphs T (5) and T (5), where this latter graph
is the Petersen graph.
The rank 3 groups have been classified by the combined effort of many people,
including Foulser, Kantor, Liebler, Liebeck and Saxl, see [212, 225, 226].
8.1.10 Strongly regular graphs from quasisymmetric de-

signs
As an application of Theorem 8.1.2, we show that quasisymmetric block designs
give rise to strongly regular graphs. A quasisymmetric design is a 2-(v, k, λ)
design (see §4.7) such that any two blocks meet in either x or y points, for
certain fixed distinct x, y. Given this situation, we may define a graph Γ on the
set of blocks, and call two blocks adjacent when they meet in x points. Let N
be the point-block matrix of the design and A the adjacency matrix of Γ. Then
N ⊤ N = kI + xA + y(J − I − A). Since each of N N ⊤ , N J, and JN is a linear
combination of I and J, we see that A2 can be expressed in terms of A, I, J,
so that Γ is strongly regular by part (ii) of Theorem 8.1.2. (For an application,
see §9.3.2.)
A large class of quasisymmetric block designs is provided by the 2-(v, k, λ)
designs with λ = 1 (also known as Steiner systems S(2, k, v)). Such designs have
only two intersection numbers since no two blocks can meet in more than one
point. This leads to a substantial family of strongly regular graphs, including
the triangular graphs T (m) (derived from the trivial design consisting of all
pairs from an m-set).
8.1.11 Symmetric 2-designs from strongly regular graphs

Conversely, some families of strongly regular graphs lead to designs. Let A be
the adjacency matrix of a strongly regular graph with parameters (v, k, λ, λ)
(i.e., with λ = µ; such a graph is sometimes called a (v, k, λ) graph). Then, by
Theorem 8.1.2
AA⊤ = A2 = (k − λ)I + λJ,
which reflects that A is the incidence matrix of a symmetric 2-(v, k, λ) design.
(And in this way one obtains precisely all symmetric 2-designs possessing a
polarity without absolute points.) For instance, the triangular graph T (6) pro-
vides a symmetric 2-(15, 8, 4) design, the complementary design of the design of
points and planes in the projective space P G(3, 2). Similarly, if A is the adja-
cency matrix of a strongly regular graph with parameters (v, k, λ, λ + 2), then
A+I is the incidence matrix of a symmetric 2-(v, k +1, λ+2) design (and in this
way one obtains precisely all symmetric 2-designs possessing a polarity with all
points absolute). For instance, the Gewirtz graph with parameters (56,10,0,2)
provides a biplane 2-(56,11,2).
8.1.12 Latin square graphs

A transversal design of strength t and index λ is a triple (X, G, B), where X is a
set of points, G is a partition of X into groups, and B is a collection of subsets
of X called blocks such that (i) t ≤ |G|, (ii) every block meets every group in
precisely one point, and (iii) every t-subset of X that meets each group in at
most one point is contained in precisely λ blocks.
Suppose X is finite and t < |G|. Then all groups G ∈ G have the same size
m, and the number of blocks is λmt . Given a point x0 ∈ X, the groups not on
x0 together with the blocks B \ {x0 } for x0 ∈ B ∈ B form a transversal design
of strength t − 1 with the same index λ.
Equivalent to the concept of transversal design is that of orthogonal array.

An orthogonal array with strength t and index λ over an alphabet of size m is a
k × N array (with N = λmt ) such that for any choice of t rows and prescribed
symbols on these rows there are precisely λ columns that satisfy the demands.
When t = 2 the strength is usually not mentioned, and one talks about
transversal designs T Dλ (k, m) or orthogonal arrays OAλ (m, k), where k is the
block size and m the group size.
When λ = 1 the index is suppressed from the notation. Now a T D(k, m) or
OA(m, k) is equivalent to a set of k − 2 mutually orthogonal Latin squares of
order m. (The k rows of the orthogonal array correspond to row index, column
index, and Latin square number; the columns correspond to the m2 positions.)
The dual of a transversal design is a net. An (m, k)-net is a set of m2 points
together with km lines, partitioned into k parallel classes, where two lines from
different parallel classes meet in precisely one point.
Given a point-line incidence structure, the point graph or collinearity graph is
the graph with the points as vertices, adjacent when they are collinear. Dually,
the block graph is the graph with the lines as vertices, adjacent when they have
a point in common.
The collinearity graph of an (m, t)-net, that is, the block graph of a transversal
design T D(t, m) (note the new use of t here!), is strongly regular with parameters
v = m2 , k = t(m − 1), λ = m − 2 + (t − 1)(t − 2), µ = t(t − 1) and eigenvalues
r = m − t, s = −t. One says that a strongly regular graph ‘is a pseudo Latin
square graph’, or ‘has Latin square parameters’ when there are t and m such
that (v, k, λ, µ) have the above values. One also says that it has ‘OA(m, t)
parameters’.
There is extensive literature on nets and transversal designs.
Proposition 8.1.7 Suppose Γ is a strongly regular graph with OA(m, t) param-

eters with a partition into cocliques of size m. Then the graph ∆ obtained from
Γ by adding edges so that these cocliques become cliques is again strongly regular
and has OA(m, t + 1) parameters.
Proof. More generally, let Γ be a strongly regular graph with a partition into
cocliques that meet the Hoffman bound. Then the graph ∆ obtained from Γ by
adding edges so that these cocliques become cliques has spectrum k + m − 1,
(r − 1)f , (s + m − 1)h , (s − 1)g−h , where m is the size of the cocliques, and
h = v/m − 1. The proposition is the special case m = r − s.
For example, from the Hall-Janko graph with OA(10, 4) parameters (100, 36,
12, 14) and a partition into ten 10-cocliques (which exists) one obtains a strongly
regular graph with OA(10, 5) parameters (100, 45, 20, 20), and hence also a sym-
metric design 2-(100,45,20). But an OA(10, 5) (three mutually orthogonal Latin
squares of order 10) is unknown.
8.1.13 Partial Geometries

A partial geometry with parameters (s, t, α) is a point-line geometry (any two
points are on at most one line) such that all lines have size s + 1, there are t + 1
lines on each point, and given a line and a point outside, the point is collinear
8.2. STRONGLY REGULAR GRAPHS WITH EIGENVALUE −2 109
with α points on the given line. One calls this structure a pg(s, t, α). Note
that the dual of a pg(s, t, α) is a pg(t, s, α) (where ‘dual’ means that the names
‘point’ and ‘line’ are swapped).
Partial geometries were introduced by Bose [31].
One immediately computes the number of points v = (s + 1)(st + α)/α and
lines b = (t + 1)(st + α)/α.
The extreme examples of partial geometries are generalized quadrangles (par-
tial geometries with α = 1) and Steiner systems S(2, K, V ) (partial geometries
with α = s + 1). Many examples are also provided by nets (with t = α) or their
duals, the transversal designs (with s = α).
The collinearity graph of a pg(s, t, α) is complete if α = s + 1, and otherwise
strongly regular with parameters v = (s + 1)(st + α)/α, k = s(t + 1), λ =
s − 1 + t(α − 1), µ = α(t + 1), and eigenvalues θ1 = s − α, θ2 = −t − 1. (Note:
earlier we used s for the smallest eigenvalue, but here s has a different meaning!)
A strongly regular graph is called geometric when it is the collinearity graph
of a partial geometry. It is called pseudo-geometric when there are integers
s, t, α such that the parameters (v, k, λ, µ) have the above-given values.
Bose [31] showed that a pseudo-geometric graph with given t and sufficiently
large s must be geometric. Neumaier [249] showed that the same conclusion
works in all cases, and hence derives a contradiction in the non-pseudo-geometric
case.
Theorem 8.1.8 (Bose-Neumaier) A strongly regular graph with s < −1 and

r > 12 s(s + 1)(µ + 1) − 1 is the block graph of a linear space or transversal
design.
It follows immediately (from this and the µ-bound) that
Theorem 8.1.9 For any fixed s = −m there are only finitely many primitive
strongly regular graphs with smallest eigenvalue s, that are not the block graph
of a linear space or transversal design with t + 1 = −m.
8.2 Strongly regular graphs with eigenvalue −2

For later use we give Seidel’s classification [274] of the strongly regular graphs
with s = −2.
Theorem 8.2.1 Let Γ be a strongly regular graph with smallest eigenvalue −2.
Then Γ is one of
(i) the complete n-partite graph Kn×2 , with parameters (v, k, λ, µ) = (2n,
2n − 2, 2n − 4, 2n − 2), n ≥ 2,
(ii) the lattice graph L2 (n) = Kn × Kn , with parameters (v, k, λ, µ) = (n2 ,
2(n − 1), n − 2, 2), n ≥ 3,
(iii) the Shrikhande graph, with parameters (v, k, λ, µ) = (16, 6, 2, 2),

(iv) the triangular graph T (n) with parameters (v, k, λ, µ) = ( n2 , 2(n − 2),
n − 2, 4), n ≥ 5,
(v) one of the three Chang graphs, with parameters (v, k, λ, µ) = (28, 12, 6, 4),
(vi) the Petersen graph, with parameters (v, k, λ, µ) = (10, 3, 0, 1),
(vii) the Clebsch graph, with parameters (v, k, λ, µ) = (16, 10, 6, 6),
(viii) the Schläfli graph, with parameters (v, k, λ, µ) = (27, 16, 10, 8).
Proof. If Γ is imprimitive, then we have case (i). Otherwise, the µ-bound gives
µ ≤ 8, and the rationality conditions give (r +2)|(µ−2)(µ−4) and integrality of
v gives µ|2r(r + 1). For µ = 2 we find the parameters of L2 (n), for µ = 4 those
of T (n), and for the remaining values for µ only the parameter sets (v, k, λ, µ) =
(10, 3, 0, 1), (16, 10, 6, 6), and (27, 16, 10, 8) survive the parameter conditions and
the absolute bound. It remains to show that the graph is uniquely determined
by its parameters in each case. Now Shrikhande [283] proved uniqueness of the
graph with L2 (n) parameters, with the single exception of n = 4, where there
is one more graph, now known as the Shrikhande graph, and Chang [79, 80]
proved uniqueness of the graph with T (n) parameters, with the single exception
of n = 8, where there are three more graphs, now known as the Chang graphs.
In the remaining three cases uniqueness is easy to see.
Let us give definitions for the graphs involved.
The Shrikhande graph is the result of Seidel switching the lattice graph L2 (4)
with respect to an induced circuit of length 8. It is the complement of the Latin
square graph for the cyclic Latin square of order 4. It is locally a hexagon.
As As As As
As As As As
Drawn on a torus: As As As As .
As As As As
The three Chang graphs are the result of switching T (8) (the line graph of K8 )
with respect to (a) a 4-coclique K4 , that is, 4 pairwise disjoint edges in K8 ; (b)
K3 + K5 , that is, 8 edges forming a triangle and a (disjoint) pentagon in K8 ;
(c) the line graph of the cubic graph formed by an 8-circuit plus edges between
opposite vertices.
The Clebsch graph is the complement of the folded 5-cube.
The Schläfli graph is the collinearity graph of GQ(2, 4) (cf. §8.6).
8.3 Connectivity
For a graph Γ, let Γi (x) denote the set of vertices at distance i from x in Γ.
Instead of Γ1 (x) we write Γ(x).
Proposition 8.3.1 If Γ is a primitive strongly regular graph, then for each

vertex x the subgraph Γ2 (x) is connected.
Proof. Note that Γ2 (x) is regular of valency k − µ. If it is not connected, then

its eigenvalue k −µ would have multiplicity at least two, and hence would be not
larger than the second largest eigenvalue r of Γ. Then x2 + (µ − λ)x + µ − k ≤ 0
for x = k − µ, i.e., (k − µ)(k − λ − 1) ≤ 0, contradiction.
The vertex connectivity κ(Γ) of a connected non-complete graph Γ is the smallest
integer m such that Γ can be disconnected by removing m vertices.
8.3. CONNECTIVITY 111
Theorem 8.3.2 ([57]). Let Γ be a connected strongly regular graph of valency

k. Then κ(Γ) = k, and the only disconnecting sets of size k are the sets of all
neighbors of some vertex x.
Proof. Clearly, κ(Γ) ≤ k. Let S be a disconnecting set of vertices not con-
taining all neighbors of some vertex. Let Γ \ S = A + B be a separation of Γ \ S.
Since the eigenvalues of A ∪ B interlace those of Γ, it follows that at least one
of A and B, say B, has largest eigenvalue at most r. It follows that the average
valency of B is at most r. Since B has an edge, r > 0.
Now let |S| ≤ k. Since B has average valency at most r, we can find two
points x, y in B such that |S ∩ Γ(x)| + |S ∩ Γ(y)| ≥ 2(k − r), so that these points
have at least k − 2r common neighbors in S.
If Γ has nonintegral eigenvalues,
√ the we have (v, k, λ, µ) = (4t + 1, 2t, t − 1, t)
for some t, and r = (−1 + v)/2. The inequality max(λ, µ) ≥ k − 2r gives
t ≤ 2, but for t = 2 the eigenvalues are integral, so we have t = 1 and Γ is the
pentagon. But the claim is true in that case.
Now let r, s be integral. If s ≤ −3, then µ = k + rs ≤ k − 3r and λ =
µ + r + s ≤ k − 2r − 3, so that no two points can have k − 2r common neighbors.
Therefore s = −2, and we have one of the eight cases in Seidel’s classification.
But not case (i), since r > 0.
Since both A and B contain an edge, both B and A have size at most
µ̄ = v − 2k + λ, so that both A and B have size at least k − λ, and v ≥ 3k − 2λ.
This eliminates cases (vii) and (viii).
If B is a clique, then |B| ≤ r + 1 = k − λ − 1, contradiction. So, B
contains two nonadjacent vertices, and their neighbors must be in B ∪ S, so
2k − µ ≤ |B| + |S| − 2 and k − µ + 2 ≤ |B| ≤ µ̄.
In cases (iii), (v), (vi) we have µ̄ = k − µ + 2, so equality holds and |B| = µ̄
and |S| = k. Since v < 2µ̄ + k, we have |A| < µ̄ and A must be a clique (of
size v − k − µ̄ = k − λ). But the Petersen graph does not contain a 3-clique,
and the Shrikhande graph does not contain 4-cliques; also, if A is a 6-clique in
a Chang graph, and a, b, c ∈ A, then Γ(a) ∩ S, Γ(b) ∩ S, and Γ(c) ∩ S are three
7-sets in the 12-set S that pairwise meet in precisely two points, impossible.
This eliminates cases (iii), (v), (vi).
We are left with the two infinite families of lattice graphs and triangular
graphs. In both cases it is easy to see that if x, y are nonadjacent, then there
exist k paths joining x and y, vertex disjoint apart from x, y, and entirely
contained in {x, y} ∪ Γ(x) ∪ Γ(y). Hence |S| = k, and if S separates x, y then
S ⊆ Γ(x) ∪ Γ(y).
The subgraph ∆ := Γ \ ({x, y} ∪ Γ(x) ∪ Γ(y)) is connected (if one removes a
point and its neighbors from a lattice graph, the result is a smaller lattice graph,
and the same
holds for a triangular graph), except in the case of the triangular
graph 52 where ∆ is empty.
Each vertex of (Γ(x) ∪ Γ(y)) \ S has a neighbor in ∆ and we find a path of
length 4 disjoint from S joining x and y, except in the case of the triangular
graph 52 , where each vertex of Γ(x) \ S is adjacent to each vertex of Γ(y) \ S,
and we find a path of length 3 disjoint from S joining x and y.
We remark that it is not true that for every strongly regular graph Γ with
vertex x the vertex connectivity of the subgraph Γ2 (x) equals its valency k − µ.
A counterexample is given by the graph Γ that is the complement of the strongly
regular graph ∆ with parameters (96,19,2,4) constructed by Haemers for q = 4,
see [175], p. 76 or [56], §8A. Indeed, we have ∆(x) ∼

= K3 + 4C4 , so that Γ2 (x)
has degree 16 and vertex connectivity 15.
Erdős & Chvátal [88] showed that if a graph Γ on at least 3 vertices has vertex
connectivity κ and largest independent set of size α, and α ≤ κ then Γ has a
Hamiltonian circuit. Bigalke & Jung [25] showed that if Γ is 1-tough, with α ≤
κ + 1 and κ ≥ 3, and Γ is not the Petersen graph, then Γ is Hamiltonian. Such
results imply for example that if Γ is strongly regular with smallest eigenvalue s,
and s is not integral, or −s ≤ µ + 1, then Γ is Hamiltonian. This, together with
explicit inspection of the Hoffman-Singleton graph, the Gewirtz graph, and the
M22 graph, shows that all connected strongly regular graphs on fewer than 99
vertices are Hamiltonian, except for the Petersen graph.
8.4 Cocliques and colorings

In §2.5 we have derived some bounds the size of a coclique in terms of eigen-
values. These bounds are especially useful for strongly regular graphs. More-
over, strongly regular graphs for which the bounds of Hoffman and Cvetković
are tight have a very special structure:
Theorem 8.4.1 Let Γ be a strongly regular graph with eigenvalues k (degree),

r and s (r > s) and multiplicities 1, f and g, respectively. Suppose that Γ is
not complete multi-partite (i.e. r 6= 0) and let C be a coclique in Γ. Then
(i) |C| ≤ g,
(ii) |C| ≤ ns/(s − k),
(iii) if |C| = g = ns/(s − k), then the subgraph Γ′ of Γ induced by the vertices
which are not in C, is strongly regular with eigenvalues k ′ = k +s (degree),
r′ = r and s′ = r + s and respective multiplicities 1, f − g + 1 and g − 1.
Proof. Parts (i) and (ii) follow from Theorems 4.1.1 and 4.1.2. Assume |C| =
g = ns/(s − k). By Theorem 2.5.4, Γ′ is regular of degree k + s. Apply
Lemma 2.11.1 to P = A − k−r n J, where A is the adjacency matrix of Γ. Since
Γ is regular, A and J commute and therefore P has eigenvalues r and s with
multiplicities f + 1 and g, respectively. We take Q = − k−r n J of size |C| = g
and R = A′ − k−r n J, where A′
is the adjacency matrix of Γ′ . Lemma 2.11.1
gives the eigenvalues of R: r (f + 1 − g times), s (0 times), r + s (g − 1 times)
and r + s + g(k − r)/n (1 time). Since Γ′ is regular of degree k + s and A′
commutes with J we obtain the required eigenvalues for A′ . By Theorem 8.1.2
Γ′ is strongly regular.
For instance, an (m−1)-coclique in T (m) is tight for both bounds and the graph
on the remaining vertices is T (m − 1).
Also for the chromatic number we can say more in case of a strongly regular
graph.
Theorem 8.4.2 If Γ is a primitive strongly regular graph, not the pentagon,

then
s
χ(Γ) ≥ 1 − .
r
8.5. AUTOMORPHISMS 113
Proof. Since Γ is primitive, r > 0 and by Corollary 4.2.4, it suffices to show

that the multiplicity g of s satisfies g ≥ −s/r for all primitive strongly regular
graphs but the pentagon. First we check this claim for all feasible parameter
sets with at most 23 vertices. Next we consider strongly regular graphs with
v ≥ 24 and r < 2. The complements of these graphs have s > −3, and by
Theorem 8.1.3 (iv), s = −2. By use of Theorem 8.2 we easily find that all these
graphs satisfy the claim.
Assume that Γ is primitive, that r ≥ 2, and that the claim does not hold
(that is g < −s/r). Now (v − 1 − g)r + gs + k = 0 gives
g 2 < −sg/r = v − 1 − g + k/r ≤ v − 1 − g + k/2 < 3v/2 − g.

p
This implies g(g + 3) ≤ 3v/2
p = 2 3v/2. By use of the absolute bound v ≤
g(g + 3)/2, we get v/2 < 2 3v/2, so v < 24. Contradiction.
For example if Γ is the complement of the triangular graph T (m) then Γ is
strongly regular with eigenvalues k = 21 (m − 2)(m − 3), r = 1 and s = 3 − m (for
m ≥ 4). The above bound gives χ(Γ) ≥ m − 2, which is tight, whilst Hoffman’s
lower bound (Theorem 4.2.2) equals 21 m. On the other hand, if m is even,
Hoffman’s bound is tight for the complement of Γ whilst the above bound is
much less. We saw (see §4.2) that a Hoffman coloring (i.e. a coloring with 1−k/s
classes) corresponds to an equitable partition of the adjacency matrix. For the
complement this gives an equitable partition into maximal cliques, which is
called a spread of the strongly regular graph. For more application of eigenvalues
to the chromatic number we refer to [140] and [158]. See also §8.7.
8.5 Automorphisms
Let A be the adjacency matrix of a graph Γ, and P the permutation matrix that
describes an automorphism φ of Γ. Then AP = P A. If φ has order m, then
P m = I, so that the eigenvalues of AP are m-th roots of unity times eigenvalues
of A.
Apply this in the special case of strongly regular graphs. Suppose φ has
f fixed points, and moves g points to a neighbor. Then f = tr P and g =
tr AP . Now consider M = A − sI. It has spectrum (k − s)1 , (r − s)f , 0g with
multiplicities written as exponents. Hence M P has eigenvalues k − s, (r − s)ζ
for certain m-th roots of unity ζ, and 0. It follows that g − sf = tr M P ≡
k − s (mod r − s).
For example, for the Petersen graph every automorphism satisfies f ≡ g +
1 (mod 3).
For example, for a hypothetical Moore graph on 3250 vertices (cf. §10.5.1),
every automorphism satisfies 8f + g ≡ 5 (mod 15).
In some cases, where a structure is given locally, it must either be a universal
object, or a quotient, where the quotient map preserves local structure, that is,
only identifies points that are far apart. In the finite case arguments like those in
this section can be used to show that f = g = 0 is impossible, so that nontrivial
quotients do not exist. For an example, see [47].
8.6 Generalized quadrangles

A generalized n-gon is a connected bipartite graph of diameter n and girth 2n.
(The girth of a graph is the length of a shortest circuit.)
It is common to call the vertices in one color class of the unique 2-coloring
points, and the other vertices lines. For example, a generalized 3-gon is the
same thing as a projective plane: any two points have an even distance at most
3, hence are joined by a line, and similarly any two lines meet in a point; finally
two lines cannot meet in two points since that would yield a quadrangle, but
the girth is 6.
A generalized quadrangle is a generalized 4-gon. In terms of points and lines,
the definition becomes: a generalized quadrangle is an incidence structure (P, L)
with set of points P and set of lines L, such that two lines meet in at most one
point, and if p is a point not on the line m, then there is a unique point q on m
and a unique line n on p such that q is on n.
8.6.1 Parameters
A generalized n-gon is called firm (thick) when each vertex has at least 2 (resp.
3) neighbors, that is, when each point is on at least two (three) lines, and each
line is on at least two (three) points.
An example of a non-firm generalized quadrangle is a pencil of lines on
one common point x0 . Each point different from x0 is on a unique line, and
Γ3 (x0 ) = ∅.
Proposition 8.6.1 (i) If a generalized n-gon Γ has a pair of opposite vertices
x, y where x has degree at least two, then every vertex has an opposite, and Γ is
firm.
(ii) A thick generalized n-gon has parameters: each line has the same number
of points, and each point is on the same number of lines. When moreover n is
odd then the number of points on each line equals the number of lines through
each point.
Proof. For a vertex x of a generalized n-gon, let k(x) be its degree. Call two
vertices of a generalized n-gon opposite when they have distance n. If x and y
are opposite then each neighbor of one is on a unique shortest path to the other,
and we find k(x) = k(y).
(i) Being non-opposite gives a bijection between Γ(x) and Γ(y), and hence if
k(x) > 1 then also each neighbor z of x has an opposite and satisfies k(z) > 1.
Since Γ is connected, it is firm.
(ii) Let x, z be two points joined by the line y. Let w be opposite to y.
Since k(w) > 2 there is a neighbor u of w opposite to both x and z. Now
k(x) = k(u) = k(z). Since Γ is connected and bipartite this shows that k(p) is
independent of the point p. If n is odd, then a vertex opposite a point is a line.

A firm, non-thick generalized quadrangle is the vertex-edge incidence graph
of a complete bipartite graph.
The halved graph of a bipartite graph Γ, is the graph on the same vertex
set, where two vertices are adjacent when they have distance 2 in Γ. The point
graph and line graph of a generalized n-gon are the two components of its halved
graph containing the points and lines, respectively.
8.6. GENERALIZED QUADRANGLES 115
The point graph and line graph of a finite thick generalized n-gon are
distance-regular of diameter ⌊n/2⌋ (see Chapter 11). In particular, the point
graph and line graph of a thick generalized quadrangle are strongly regular (see
Theorem 8.6.2).
It is customary to let GQ(s, t) denote a finite generalized quadrangle with
s + 1 points on each line and t + 1 lines on each point. Note that it is also
customary to use s to denote the smallest eigenvalue of a strongly regular graph,
so in this context one has to be careful to avoid confusion.
It is a famous open problem whether a thick generalized n-gon can have
finite s and infinite t. In the special case of generalized quadrangles a little
is known: Cameron, Kantor, Brouwer, and Cherlin [74, 44, 81] show that this
cannot happen for s + 1 ≤ 5.
8.6.2 Constructions of generalized quadrangles

Suppose V is a vector space provided with a nondegenerate quadratic form f
of Witt index 2 (that is, such that the maximal totally singular subspaces have
vector space dimension 2). Consider in the projective space P V the singular
projective points and the totally singular projective lines. These will form a
generalized quadrangle.
Indeed, f defines a bilinear form B on V via B(x, y) = f (x+y)−f (x)−f (y).
Call x and y orthogonal when B(x, y) = 0. When two singular vectors are
orthogonal, the subspace spanned by them is totally singular. And conversely,
in a totally singular subspace any two vectors are orthogonal. The collection of
all vectors orthogonal to a given vector is a hyperplane. We have to check that
if P = hxi is a singular projective point, and L is a totally singular projective
line not containing P , then P has a unique neighbor on L. But the hyperplane
of vectors orthogonal to x meets L, and cannot contain L otherwise f would
have larger Witt index.
This construction produces generalized quadrangles over arbitrary fields. If

V is a vector space over a finite field Fq , then a nondegenerate quadratic
form can have Witt index 2 in dimensions 4, 5, and 6. A hyperbolic quadric
in 4 dimensions yields a generalized quadrangle with parameters GQ(q, 1), a
parabolic quadric in 5 dimensions yields a generalized quadrangle with parame-
ters GQ(q, q), and an elliptic quadric in 6 dimensions yields a generalized quad-
rangle with parameters GQ(q, q 2 ).
Other constructions, and other parameters occur.
In the below we’ll meet GQ(2, t) for t = 1, 2, 4 and GQ(3, 9). Let us give simple
direct descriptions for GQ(2, 1) and GQ(2, 2).
The unique GQ(2, 1) is the 3-by-3 grid: 9 points, 6 lines. Its point graph is
K3 × K3 .
The unique GQ(2, 2) is obtained by taking as points the 15 pairs from a 6-
set, and as lines the 15 partitions of that 6-set into three pairs. Now collinearity
is being disjoint. Given a point ac, and a line {ab, cd, ef }, the two points ab
and cd on this line are not disjoint from ac, so that ef is the unique point on
this line collinear with ac, and the line joining ac and ef is {ac, bd, ef }.
8.6.3 Strongly regular graphs from generalized quadran-

gles
As mentioned before, the point graph (collinearity graph) of a finite thick gen-
eralized quadrangle is strongly regular. The parameters and eigenvalue can be
obtained in a straightforward way (see exercises).
Theorem 8.6.2 The collinearity graph of a finite generalized quadrangle with

parameters GQ(s, t) is strongly regular with parameters
v = (s + 1)(st + 1), k = s(t + 1), λ = s − 1, µ = t + 1
and spectrum
s(t + 1) with multiplicity 1,

s−1 with multiplicity st(s + 1)(t + 1)/(s + t),
−t − 1 with multiplicity s2 (st + 1)/(s + t).
In particular, if a GQ(s, t) exists, then (s + t)|s2 (st + 1).
8.6.4 Generalized quadrangles with lines of size 3

Let a weak generalized quadrangle be a point-line geometry with the properties
that two lines meet in at most one point, and given a line m and a point
p outside there is a unique pair (q, n) such that p ∼ n ∼ q ∼ m, where ∼
denotes incidence. The difference with the definition of a generalized quadrangle
is that connectedness is not required. (But of course, as soon as there is a point
and a line then the geometry is connected.)
Theorem 8.6.3 A weak generalized quadrangle where all lines lines have size
3 is one of the following:
(i) a coclique (no lines),

(ii) a pencil (all lines passing through a fixed point),
(iii) the unique GQ(2, 1),
(iv) the unique GQ(2, 2),
(v) the unique GQ(2, 4).
Proof. After reducing to the case of GQ(2, t) one finds (t + 2)|(8t + 4), i.e.,
(t + 2)|12, i.e., t ∈ {1, 2, 4, 10}, and t = 10 is ruled out by the Krein conditions.
Alternatively, or afterwards, notice that the point graphs have eigenvalue 1, so
that their complements have smallest eigenvalue −2, and apply Seidel’s classi-
fication. Cases (iii), (iv), (v) here have point graphs that are the complements
of the lattice graph K3 × K3 , the triangular graph T (6), and the Schläfli graph,
respectively.
This theorem can be used in the classification of root lattices, where the five
cases correspond to An , Dn , E6 , E7 , E8 , respectively.
And the classification of root lattices can be used in the classification of
graphs with smallest eigenvalue −2. Indeed, for such graphs A + 2I is posi-
tive semidefinite, and one can represent these graphs by vectors in a Euclidean
space such that (x, y) = 2, 1, 0 when x and y are equal, adjacent, nonadjacent,
respectively. The lattice spanned by these vectors is a root lattice.
8.7. THE (81,20,1,6) STRONGLY REGULAR GRAPH 117
8.7 The (81,20,1,6) strongly regular graph

Large parts of this section are taken from [51]. Sometimes the graph of this
section is called the Brouwer-Haemers graph.
Let Γ = (X, E) be a strongly regular graph with parameters (v, k, λ, µ) =

(81, 20, 1, 6). Then Γ has spectrum {201 , 260 , −720 }, where the exponents de-
note multiplicities. We will show that up to isomorphism there is a unique
such graph Γ. More generally we give a short proof for the fact (due to
Ivanov & Shpectorov [206]) that a strongly regular graph with parameters
(v, k, λ, µ) = (q 4 , (q 2 + 1)(q − 1), q − 2, q(q − 1)) that is the collinearity graph
of a partial quadrangle (that is, in which all maximal cliques have size q) is
the second subconstituent of the collinearity graph of a generalized quadrangle
GQ(q, q 2 ). In the special case q = 3 this will imply our previous claim, since
λ = 1 implies that all maximal cliques have size 3, and it is known (see Cameron,
Goethals & Seidel [76]) that there is a unique generalized quadrangle GQ(3, 9)
(and this generalized quadrangle has an automorphism group transitive on the
points).
8.7.1 Descriptions
Let us first give a few descriptions of our graph on 81 vertices. Note that the
uniqueness shows that all constructions below give isomorphic graphs, some-
thing, which is not immediately obvious from the description in all cases.
A. Let X be the point set of AG(4, 3), the 4-dimensional affine space over F3 ,
and join two points when the line connecting them hits the hyperplane at infinity
(a P G(3, 3)) in a fixed elliptic quadric Q. This description shows immediately
that v = 81 and k = 20 (since |Q| = 10). Also λ = 1 since no line meets Q
in more than two points, so that the affine lines are the only triangles. Finally
µ = 6, since a point outside Q in P G(3, 3) lies on 4 tangents, 3 secants and 6
exterior lines with respect to Q, and each secant contributes 2 to µ. We find that
the group of automorphisms contains G = 34 · P GO4− · 2, where the last factor 2
accounts for the linear transformations that do not preserve the quadratic form
Q, but multiply it by a constant. In fact this is the full group, as will be clear
from the uniqueness proof.
B. A more symmetric form of this construction is found by starting with X =

1⊥ / h1i in F63 provided with the standard bilinear form. The corresponding
quadratic form (Q(x) = wt(x), the number of nonzero coordinates of x) is
elliptic, and if we join two vertices x + h1i , y + h1i of X when Q(x − y) = 0,
i.e., when their difference has weight 3, we find the same graph as under A.
This construction shows that the automorphism group contains G = 34 · (2 ×
Sym (6)) · 2, and again this is the full group.
C. There is a unique strongly regular graph Γ with parameters (112, 30, 2, 10),
the collinearity graph of the unique generalized quadrangle with parameters
GQ(3, 9). Its second subconstituent is an (81, 20, 1, 6) strongly regular graph,
and hence isomorphic to our graph Γ. (See Cameron, Goethals & Seidel [76].)
We find that Aut Γ contains (and in fact it equals) the point stabilizer in U4 (3) ·
D8 acting on GQ(3, 9).
D. The graph Γ is the coset graph of the truncated ternary Golay code C: take
the 34 cosets of C and join two cosets when they contain vectors differing in
only one place.
E. The graph Γ is the Hermitean forms graph on F29 ; more generally, take
the q 4 matrices M over Fq2 satisfying M ⊤ = M , where − denotes the field
automorphism x → xq (applied entrywise), and join two matrices when their
difference has rank 1. This will give us a strongly regular graph with parameters
(v, k, λ, µ) = (q 4 , (q 2 + 1)(q − 1), q − 2, q(q − 1)).
F. The graph Γ is the graph with vertex set F81 , where two vertices are joined
when their difference is a fourth power. (This construction was given by Van
Lint & Schrijver [228].)
8.7.2 Uniqueness
Now let us embark upon the uniqueness proof. Let Γ = (X, E) be a strongly
regular graph with parameters (v, k, λ, µ) = (q 4 , (q 2 + 1)(q − 1), q − 2, q(q − 1))
and assume that all maximal cliques (we shall just call them lines) of Γ have
size q. Let Γ have adjacency matrix A. Using the spectrum of A - it is {k 1 , (q −
1)f , (q − 1 − q 2 )g }, where f = q(q − 1)(q 2 + 1) and g = (q − 1)(q 2 + 1) - we can
obtain some structure information. Let T be the collection of subsets of X of
cardinality q 3 inducing a subgraph that is regular of degree q − 1.
1. Claim. If T ∈ T, then each point of X \ T is adjacent to q 2 points of T .
Look at the matrix B of average row sums of A, with sets of rows and
columns partitioned according to {T, X \ T }. We have

q − 1 q 2 (q − 1)
B=
q2 k − q2
with eigenvalues k and q − 1 − q 2 , so interlacing is tight, and by Corol-

lary 2.5.4(ii) it follows that the row sums are constant in each block of
A.
2. Claim. Given a line L, there is a unique TL ∈ T containing L.
Let Z be the set of vertices in X \ L without a neighbor in L. Then
|Z| = q 4 − q − q(k − q + 1) = q 3 − q. Let T = L ∪ Z. Each vertex of Z is
adjacent to qµ = q 2 (q − 1) vertices with a neighbor in L, so T induces a
subgraph that is regular of degree q − 1.
3. Claim. If T ∈ T and x ∈ X \ T , then x is on at least one line L disjoint
from T , and TL is disjoint from T for any such line L.
The point x is on q 2 + 1 lines, but has only q 2 neighbors in T . Each point
of L has q 2 neighbors in T , so each point of T has a neighbor on L and
hence is not in TL .
4. Claim. Any T ∈ T induces a subgraph ∆ isomorphic to q 2 Kq .
It suffices to show that the multiplicity m of the eigenvalue q −1 of ∆ is (at
least) q 2 (it cannot be more). By interlacing we find m ≥ q 2 −q, so we need
some additional work. Let M := A − (q − 1/q 2 )J. Then M has spectrum
{(q −1)f +1 , (q −1−q 2 )g }, and we want that MT , the submatrix of M with
rows and columns indexed by T , has eigenvalue q − 1 with multiplicity (at
8.7. THE (81,20,1,6) STRONGLY REGULAR GRAPH 119
least) q 2 −1, or, equivalently (by Lemma 2.11.1), that MX\T has eigenvalue
q − 1 − q 2 with multiplicity (at least) q − 2. But for each U ∈ T with
U ∩ T = ∅ we find an eigenvector xU = (2 − q)χU + χX\(T ∪U ) of MX\T
with eigenvalue q − 1 − q 2 . A collection {xU |U ∈ U} of such eigenvectors
cannot be linearly
S dependent
S when U = {U1 , U2 , . . .} can be ordered such
that Ui 6⊂ j<i Uj and U 6= X \ T , so we can find (using Claim 3) at
least q − 2 linearly independent such eigenvectors, and we are done.
5. Claim. Any T ∈ T determines a unique partition of X into members of
T.
Indeed, we saw this in the proof of the previous step.
Let Π be the collection of partitions of X into members of T. We have |T| =
q(q 2 + 1) and |Π| = q 2 + 1. Construct a generalized quadrangle GQ(q, q 2 ) with
point set {∞} ∪ T ∪ X as follows: The q 2 + 1 lines on ∞ are {∞} ∪ π for π ∈ Π.
The q 2 remaining lines on each T ∈ T are {T } ∪ L for L ⊂ T . It is completely
straightforward to check that we really have a generalized quadrangle GQ(q, q 2 ).
8.7.3 Independence and chromatic numbers

We have α(Γ) = 15 and χ(Γ) = 7.
Clearly, the independence number of our graph is one less than the indepen-
dence number of the unique GQ(3, 9) of which it is the second subconstituent.
So it suffices to show that α(∆) = 16, where ∆ is the collinearity graph of
GQ(3, 9).
It is easy to indicate a 16-coclique: define GQ(3, 9) in P G(5, 3) provided
P6
with the nondegenerate elliptic quadratic form i=1 x2i . There are 112 isotropic
points, 80 of weight 3 and 32 of weight 6. Among the 32 of weight 6, 16 have
coordinate product 1, and 16 have coordinate product −1, and these two 16-sets
are cocliques.
That there is no larger coclique can be seen by cubic counting.
Let C be a 16-coclique in ∆. Let there be ni vertices outside that have i
neighbors inside. Then
X X X i X i
ni = 96, ini = 480, ni = 1200, ni = 2240,
2 3
so that X
(i − 4)2 (i − 10)ni = 0.
(Here the quadratic counting is always possible in a strongly regular graph,
and the last equation can be written because the second subconstituent is itself
strongly regular.) Now each point is on 10 lines, and hence cannot have more
than 10 neighbors in C. It follows that each point has either 4 or 10 neighbors
in C. In particular, C is maximal.
As an aside: Solving these equations gives n4 = 80, n10 = 16. Let D be the
set of 16 vertices with 10 neighbors in C. If two vertices d1 , d2 ∈ D are adjacent
then they can have only 2 common neighbors in C, but each has 10 neighbors
in C, contradiction. So, also D is a 16-coclique, which means that 16-cocliques
in ∆ come in pairs.
Since 81/15 > 5, we have χ(Γ) ≥ 6. Since ∆ has a split into two Gewirtz graphs,
and the Gewirtz graph has chromatic number 4, it follows that χ(∆) ≤ 8. (And
in fact equality holds.) This shows that for our graph 6 ≤ χ(Γ) ≤ 8. In fact
χ(Γ) = 7 can be seen by computer (Edwin van Dam, pers. comm.).
Since λ = 1, the maximum clique size equals 3. And from the uniqueness
proof it is clear that Γ admits a partition into 27 triangles. So the complement
of Γ has chromatic number 27.
8.8 Strongly regular graphs and 2-weight codes

8.8.1 Codes, graphs and projective sets
In this section we show the equivalence of three kinds of objects:
(i) projective two-weight codes,
(ii) subsets X of a projective space such that |X ∩ H| takes two values when
H ranges through the hyperplanes of the projective space,
(iii) strongly regular graphs defined by a difference set that is a cone in a vector
space.
This equivalence is due to Delsarte [126]. An extensive survey of this material

was given by Calderbank & Kantor [73].
A linear code is a linear subspace of some finite vector space with fixed
basis. For basic terminology and results on codes, see MacWilliams & Sloane
[236] and Van Lint [227]. A linear code C is called projective when its dual C ⊥
has minimum weight at least three, that is, when no two coordinate positions
of C are linearly dependent. The weight of a vector is its number of nonzero
coordinates. A two-weight code is a linear code in which precisely two nonzero
weights occur.
Let us first discuss the correspondence between linear codes and subsets of
projective spaces.
8.8.2 The correspondence between linear codes and sub-

sets of a projective space
A linear code C of word length n over the alphabet Fq is a linear subspace of the
vector space Fnq . The weight of a vector is its number of nonzero coordinates.
We call C an [n, m, w]-code if C has dimension m and minimum nonzero weight
w. We say that C has effective length (or support) n − z when there are precisely
z coordinate positions j such that cj = 0 for all c ∈ C. The dual C ⊥ of Pa code
C is the linear code {d ∈ Fnq | hc, di = 0 for all u ∈ C}, where hc, di = ci di is
the standard inner product (bilinear form).
Let us call two linear codes of length n over Fq equivalent when one arises
from the other by permutation of coordinates or multiplication
of coordinates
by

1111 1212
a nonzero constant. E.g., the F3 -codes generated by and
0012 1100
are equivalent. If we study codes up to equivalence, and assume that n is chosen
minimal, i.e., that the generator matrix has no zero columns, we may identify
the set of columns in a m × n generator matrix with points of a projective space
8.8. STRONGLY REGULAR GRAPHS AND 2-WEIGHT CODES 121
P G(m − 1, q). In this way, we find a subset X of P G(m − 1, q), possibly with
repeated points, or, if you prefer, a weight function w : P G(m − 1, q) → N.
Choosing one code in an equivalence class means choosing a representative
in Fm
q for each x ∈ X, and fixing an order on X. Now the code words can be
identified with the linear functionals f , and the x-coordinate position is f (x).
Clearly, the code has word length n = |X|. Note that the code will have
dimension m if and only if X spans P G(m − 1, q), i.e., if and only if X is not
contained in a hyperplane.
The weight of the code word f equals the number of x such that f (x) 6= 0.
But a nonzero f vanishes on a hyperplane of P G(m − 1, q). Consequently, the
number of words of nonzero weight w in the code equals q − 1 times the number
of hyperplanes H that meet X in n − w points. In particular the minimum
distance of the code is n minus the maximum size of H ∩ X for a hyperplane
H.
The minimum weight of the dual code equals the minimum number of points
of X that are dependent. So, it is 2 if and only if X has repeated points, and 3
when X has no repeated points but has three collinear points.
Example Take for X the entire projective space P G(m − 1, q), so that n =
|X| = (q m −1)/(q−1). We find the so-called simplex code: all words have weight
q m−1 , and we have an [n, m, q m−1 ]-code over Fq . Its dual is the [n, n − m, 3]
Hamming code. It is perfect!
8.8.3 The correspondence between projective two-weight

codes, subsets of a projective space with two inter-
section numbers, and affine strongly regular graphs
Given a subset X of size n of P G(m − 1, q), let us define a graph Γ with vertex
set Fmq , with x ∼ y if and only if hy − xi ∈ X. Then clearly Γ is regular of
valency k = (q − 1)n. We show below that this graph has eigenvalues k − qwi
when the linear code has weights wi . Hence if a linear code has only two nonzero
weights, and its dual has minimum weight at least 3, then we have a strongly
regular graph.
Let us look at the details.
Let F = Fq and K = Fqk and let tr : K → F be the trace map defined by
k−1
tr(x) = x + xq + ... + xq . Then the F -linear maps f : F → K are precisely
the maps fa defined by fa (x) = tr(ax), for a ∈ K. If tr(ax) = 0 for all x, then
a = 0.
(Indeed, first of all, these maps are indeed F -linear. If a 6= 0, then tr(ax) is
a polynomial of degree q k−1 and cannot have q k zeros. It follows that we find
q k distinct maps fa . But this is the total number of F -linear maps from K to F
(since K is a vector space of dimension k over F , and such a map is determined
by its values on a basis).)
Let G be a finite abelian group. If a : G → C is any function, and we
define the matrix A by Axy = a(y − x), then the eigenspaces of A have a basis
consisting of characters of G.
(Indeed, if χ : G → C∗ is a character (a homomorphism from the additively
written group G into the multiplicative group of nonzero complex numbers),
then !
X X
(Aχ)x = a(y − x)χ(y) = a(z)χ(z) χ(x)
y∈G z∈G
so
P that χ (regarded as column vector) is an eigenvector of A with eigenvalue
z∈G a(z)χ(z). But G has |G| characters, and these are linearly independent,
so this gives us the full spectrumof A.)   
a b c 1
Example. The matrix A =  c a b  has eigenvectors  ω  with
b c a ω2
2
eigenvalues a + bω + cω , where ω runs through the cube roots of unity.
Now apply this to the adjacency matrix A of the graph Γ. Let D := {d ∈
Fmq | hdi ∈ X}, so that |D| = (q−1).|X|. Then the neighbors of the vertex x of Γ
are the points x + d for d ∈ D, and wePsee that Γ has valency k = |D| = (q − 1)n.
The eigenvalues of A are the sums d∈D χ(d), where χ is a character of the
additive group of Fm q . Let ζ = e
2πi/p
be a primitive p-th root of unity, and let
tr : Fq → Fp be the trace function. Then the characters χ are of the form
χa (x) = ζ tr(ha,xi) .
Now
X q if ha, xi = 0
χa (λx) =
0 otherwise.
λ∈Fq
(Indeed, if S denotes this sum, then χa (µx)S = S for all µ, so if S 6= 0, then

tr(ha, µxi) = tr(µha, xi) = 0 for all µ, and by the above ha, xi = 0.)
Thus, we find, if D0 is a set of representatives for X,
X X X
χa (d) = χa (λd) = q.|Ha ∩ X| − |X|
d∈D d∈D0 λ∈Fq \{0}
where Ha is the hyperplane {hxi | ha, xi = 0} in P G(m − 1, q). This shows

that if Ha meets X in ma points, so that the corresponding q − 1 code words
have weight wa = n − ma , then the corresponding eigenvalue is qma − n =
(q − 1)n − qwa = k − qwa .
We have proved:
Theorem 8.8.1 There is a 1-1-1 correspondence between
(i) linear codes C of effective word length n and dimension m and (q − 1)fi
words of weight wi , and
(ii) weighted subsets X of total size n of the projective space P G(m − 1, q)
such that for fi hyperplanes H we have |X \ H| = wi , and
(iii) graphs Γ, without loops but possibly with multiple edges, with vertex set Fm
q ,
invariant under translation and dilatation, and with eigenvalues k − qwi
of multiplicity (q − 1)fi , where k = n(q − 1).
If the code C is projective, that is, is no two coordinate positions are de-
pendent (i.e., is the dual code has minimum weight at least 3), then X has no
repeated points, and we find an ordinary subset under (ii), and a simple graph
under (iii) (that is, without multiple edges).
Corollary 8.8.2 There is a 1-1-1 correspondence between
(i) projective linear codes C of effective word length n and dimension m with
precisely two nonzero weights w1 and w2 , and
(ii) subsets X of size n of the projective space P G(m − 1, q) such that for each
hyperplane H we have |X \ H| = wi , i ∈ {1, 2}, and
(iii) strongly regular graphs Γ, with vertex set Fm

q , invariant under translation
and dilatation, and with eigenvalues k − qwi , where k = n(q − 1).
For example, if we take a hyperoval in P G(2, q), q even, we find a two-

weight [q + 2, 3, q]-code over Fq . If we take the curve {(1, t, t2 , ..., tm−1 ) | t ∈
Fq } ∪ {(0, 0, ..., 0, 1)} in P G(m − 1, q), q arbitrary, we find a [q + 1, m, q − m + 2]-
code over Fq . (These codes are optimal: they reach the Singleton bound.)
A 1-1 correspondence between projective codes and 2-weight codes was
shown in Brouwer & van Eupen [50].
8.8.4 Duality for affine strongly regular graphs

Let X be a subset of P G(m − 1, q) such that all hyperplanes meet it in either
m1 or m2 points. In the dual projective space (where the rôles of points and
hyperplanes have been interchanged), the collection Y of hyperplanes that meet
X in m1 points, is a set with the same property: there are numbers n1 and n2
such that each point is in either n1 or in n2 hyperplanes from Y .
Indeed, let x ∈ X be in n1 hyperplanes from Y . We can find n1 (independent
of the choice of x) by counting hyperplanes on pairs x, y of distinct points in X:
q m−1 − 1 q m−2 − 1
n1 .(m1 − 1) + ( − n1 ).(m2 − 1) = (|X| − 1). .
q−1 q−1
In a similar way we find n2 , the number of hyperplanes from Y on a point
outside X. Computation yields (m1 − m2 )(n1 − n2 ) = q k−2 . This proves:
Proposition 8.8.3 The difference of the weights in a projective 2-weight code,

and the difference of the nontrivial eigenvalues of an affine strongly regular
graph, is a power of p, where p is the characteristic of the field involved.
Let Γ and ∆ be the strongly regular graphs corresponding to X and Y ,

respectively. We see that Γ and ∆ both have q k vertices; Γ has valency k = (q −
1)|X| and multiplicity f = (q − 1)|Y |, and for ∆ these values have interchanged
rôles. We call ∆ the dual of Γ. (More generally it is possible to define the dual of
an association scheme with a regular abelian group of automorphisms, cf. [48],
p. 68.))
Example. The ternary Golay code is a perfect [11, 6, 5] code over F3 ,
and its dual C is a [11, 5, 6] code with weights 6 and 9. The correspond-
ing strongly regular graph Γ has parameters (v, k, v − k − 1, λ, µ, r, s, f, g) =
(243, 22, 220, 1, 2, 4, −5, 132, 110) (it is the Berlekamp-van Lint-Seidel graph)
and its dual has parameters (243, 110, 132, 37, 60, 2, −25, 220, 22), and we see
that k, v − k − 1 interchange place with g, f . The code corresponding to ∆ is a
[55, 5, 36] ternary code.
Example. The quaternary Hill code ([193]) is a [78, 6, 56] code over F4
with weights 56 and 64. The corresponding strongly regular graph has pa-
rameters (4096, 234, 3861, 2, 14, 10, −22, 2808, 1287). Its dual has parameters
(4096, 1287, 2808, 326, 440, 7, −121, 3861, 234), corresponding to a quaternary
[429, 6, 320] code with weights 320 and 352. This code lies outside the range
of the tables, but its residue is a world record [109, 5, 80] code. The binary
[234, 12, 112] code derived from the Hill code has a [122, 11, 56] code as residue
- also this is a world record.
8.8.5 Cyclotomy
In this section we take D to be a union of cosets of a subgroup of the multiplica-
tive group of a field Fq . (I.e., the q here corresponds to the q k of the previous
sections.)
Let q = pκ , p prime and e|(q − 1), say q = em + 1. Let K ⊆ F∗q be the
subgroup of the e-th powers (so that |K| = m). Let α be a primitive S element
of Fq . For J ⊆ {0, 1, ..., e − 1} put u := |J| and D := DJ := {αj K | j ∈
J} = {αie+j |j ∈ J, 0 ≤ i < m}. Define a (directed) graph Γ = ΓJ with vertex
set Fq and edges (x, y) whenever y − x ∈ D. Note that Γ will be undirected iff
either −1 is an e-th power (i.e., q is even or e|(q − 1)/2) or J + (q − 1)/2 = J
(arithmetic in Ze ).
Let A = AJ be the adjacency matrix of Γ defined by A(x, y) = 1 if (x, y) is
an edge of Γ and = 0 otherwise. Let us compute the eigenvalues of A. For each
(additive) character χ of Fq we have
X X
(Aχ)(x) = χ(y) = ( χ(u))χ(x).
y∼x u∈D
So each character gives us an eigenvector, and since these are all independent
we know all eigenvalues. Their explicit determination requires some theory of
Gauss sums. Let us write Aχ = θ(χ)χ. Clearly, θ(1) = mu, the valency of Γ.
Now assume χ 6= 1. Then χ = χg for some g, where
2πi
χg (αj ) = exp( tr(αj+g ))
p
and tr : Fq → Fp is the trace function.

If µ is any multiplicative character of order e (say, µ(αj ) = ζ j , where ζ =
exp( 2πi
e )), then
e−1
X
e if µ(x) = 1
µi (x) =
0 otherwise.
i=0
Hence,
X XX e−1
1XX X
θ(χg ) = χg (u) = χj+g (u) = χj+g (x) µi (x) =
e ∗ i=0
u∈D j∈J u∈K j∈J x∈Fq
e−1 X e−1
1X X 1X X
= (−1 + χj+g (x)µi (x)) = (−1 + µ−i (αj+g )Gi )
e i=1
e i=1
j∈J x6=0 j∈J
P
where Gi is the Gauss sum x6=0 χ0 (x)µi (x).
In general, determination of Gauss sums seems to be complicated, but there
are a few explicit results. For our purposes the most interesting is the following:
Proposition 8.8.4 (Stickelberger and Davenport & Hasse, see McEliece &
Rumsey [242]) Suppose e > 2 and p is semiprimitive mod e, i.e., there ex-
ists an l such that pl ≡ −1 (mod e). Choose l minimal and write κ = 2lt.
Then
√
Gi = (−1)t+1 εit q,
where
−1 if e is even and (pl + 1)/e is odd
ε=
+1 otherwise.
Under the hypotheses of this proposition, we have

e−1
X e−1
X √
−i j+g −i(j+g) t+1 it √ (−1)t q if r 6= 1,
µ (α )Gi = ζ (−1) ε q= t+1 √
(−1) q(e − 1) if r = 1,
i=1 i=1
where ζ = exp(2πi/e) and r = rg,j = ζ −j−g εt (so that re = εet = 1), and hence
u √ √
θ(χg ) = (−1 + (−1)t q) + (−1)t+1 q.#{j ∈ J|rg,j = 1}.
e
If we abbreviate the cardinality in this formula with # then: If εt = 1 then
# = 1 if g ∈ −J (mod e), and = 0 otherwise. If εt = −1 (then e is even and p
is odd) then # = 1 if g ∈ 12 e − J (mod e), and = 0 otherwise. We proved:
Theorem 8.8.5 Let q = pκ , p prime and e|(q − 1), where p is semiprimitive

mod e, i.e., there is an l > 0 such that pl ≡ −1 mod e. Choose l minimal with
this property and write κ = 2lt. Choose u, 1 ≤ u ≤ e − 1 and assume that q
is even or u is even or e|(q − 1)/2. Then the graphs ΓJ (where J is arbitrary
for q even or e|(q − 1)/2 and satisfies J + (q − 1)/2 = J mod e otherwise) are
strongly regular with eigenvalues
q−1
k= e u with multiplicity 1,
u √
θ1 = + (−1)t q)
e (−1 with multiplicity q − 1 − k,
√ √
θ2 = ue (−1 + (−1)t q) + (−1)t+1 q with multiplicity k.
(Obviously, when t is even we have r = θ1 , s = θ2 , and otherwise r = θ2 ,

s = θ1 .)
Clearly, if e|e′ |(q − 1), then the set of e-th powers is a union of cosets of the
set of e′ -th powers, so when applying the above theorem we may assume that e
has been chosen as large as possible, i.e., e = pl + 1. Then the restriction ‘q is
even or u is even or e|(q −1)/2’ is empty, and J can always be chosen arbitrarily.
The above construction can be generalized. Pick several values ei (i ∈ I)
with ei |(q − 1). Let Ki be the subgroup of F∗q of the ei -th powers. Let Ji be a
S S
subset of {0, 1, ..., ei − 1}. Let Di := DJi := {αj Ki | j ∈ Ji }. Put D := Di .
If the Di are mutually disjoint, then D defines a graph of which we can compute
the spectrum.
For example, let p be odd, and take ei = pli + 1 (i = 1, 2) and q = pκ where
κ = 4li si (i = 1, 2). Pick J1 to consist of even numbers only, and J2 to consist
of odd numbers only. Then D1 ∩ D2 = ∅ and g ∈ −Ji (mod ei ) cannot happen

for i = 1, 2 simultaneously. This means that the resulting graph will be strongly
regular with eigenvalues
|J1 | |J2 | √ √
θ(χg ) = ( + )(−1 + q) − q.δ(g ∈ −Ji (mod ei ) for i = 1 or i = 2)
e1 e2
(where δ(P ) = 1 if P holds, and δ(P ) = 0 otherwise). See also [62]. In the
special case p = 3, l1 = 1, l2 = 2, e1 = 4, e2 = 10, J1 = {0}, J2 = {1}, the
difference set consists of the powers αi with i ≡ 0 (mod 4) or i ≡ 1 (mod 10), i.e.,
is the set {1, α, α4 , α8 , α11 , α12 , α16 }hα20 i, and we found the first graph from De
Lange [220] again. (It has parameters (v, k, λ, µ) = (6561, 2296, 787, 812) and
spectrum 22961 284264 (−53)2296 .)
8.9 Table
Below a table with the feasible parameters for strongly regular graphs on at most
100 vertices. Here feasible means that the parameters v, k, λ, µ and multiplicities
f, g are integers, with 0 ≤ λ < k − 1 and 0 < µ < k < v. In some cases a feasible
parameter set is ruled out by the absolute bound or the Krein conditions, or the
restriction that the order of a conference graph must be the sum of two squares.
For some explanation of the comments, see after the table.
∃ v k λ µ rf sg comments
! 5 2 0 1 0.6182 −1.6182 pentagon; Paley(5); Seidel 2-graph−∗
! 9 4 1 2 14 −24 Paley(9); 32 ; 2-graph−∗
! 10 3 0 1 15 −24 Petersen graph [259]; N O4− (2); N O3−⊥ (5); 2-graph
5
6 3 4 14 −25 2 ; 2-graph
! 13 6 2 3 1.3036 −2.3036 Paley(13); 2-graph−∗
! 15 6 1 3 19 −35 O5 (2) polar graph; Sp4 (2) polar graph; N O4− (3); 2-
graph−∗
6
8 4 4 25 −29 2 ; 2-graph−∗
! 16 5 0 2 110 −35 2
q22 = 0; vanLint-Schrijver(1); V O4− (2) affine polar
graph; projective binary [5,4] code with weights 2, 4;
RSHCD − ; 2-graph
10 6 6 25 −210 Clebsch graph [93, 99, 92, 274]; q11 1
= 0; vanLint-
Schrijver(2); 2-graph
2! 16 6 2 2 26 −29 2
Shrikhande graph [283]; 4 ; from a partial spread:
projective binary [6,4] code with weights 2, 4;
RSHCD + ; 2-graph
9 4 6 19 −36 OA(4,3); Bilin2×2 (2); Goethals-Seidel(2,3); V O4+ (2)
affine polar graph; 2-graph
! 17 8 3 4 1.5628 −2.5628 Paley(17); 2-graph−∗
! 21 10 3 6 114 −46
7
10 5 4 36 −214 2
− 21 10 4 5 1.79110 −2.79110 Conf
! 25 8 3 2 38 −216 52
16 9 12 116 −48 OA(5,4)
15! 25 12 5 6 212 −312 complete enumeration by Paulus [255]; Paley(25);
OA(5,3); 2-graph−∗
10! 26 10 3 4 213 −312 complete enumeration by Paulus [255]; 2-graph
15 8 9 212 −313 S(2,3,13); 2-graph
! 27 10 1 5 120 −56 2
q22 = 0; O6− (2) polar graph; GQ(2, 4); 2-graph−∗
16 10 8 46 −220 1
Schläfli graph; unique by Seidel [274]; q11 = 0; 2-
graph−∗
− 28 9 0 4 121 −56 Krein2 ; Absolute bound
8.9. TABLE 127
18 12 10 46 −221 Krein1 ; Absolute bound
47 −220 Chang graphs [80]; 82 ; 2-graph

4! 28 12 6 4
15 6 10 120 −57 N O6+ (2); Goethals-Seidel(3,3); Taylor 2-graph for
U3 (3)
41! 29 14 6 7 2.19314 −3.19314 complete enumeration by Bussemaker & Spence
[pers.comm.]; Paley(29); 2-graph−∗
− 33 16 7 8 2.37216 −3.37216 Conf
3854! 35 16 6 8 220 −414 complete enumeration by McKay & Spence [244]; 2-
graph−∗
18 9 9 314 −320 S(2,3,15); lines in P G(3, 2); O6+ (2) polar graph; 2-
graph−∗
! 36 10 4 2 410 −225 62
25 16 20 125 −510
180! 36 14 4 6 221 −414 U3 (3).2/L2 (7).2 - subconstituent of the Hall-Janko
graph; complete enumeration by McKay & Spence
[244]; RSHCD − ; 2-graph
21 12 12 314 −321 2-graph
9
58 −227

! 36 14 7 4 2
21 10 15 127 −68
32548! 36 15 6 6 315 −320 complete enumeration by McKay & Spence [244];
OA(6,3); N O6− (2); RSHCD + ; 2-graph
20 10 12 220 −415 N O5− (3); 2-graph
+ 37 18 8 9 2.54118 −3.54118 Paley(37); 2-graph−∗
28! 40 12 2 4 224 −415 complete enumeration by Spence [286]; O5 (3) polar
graph; Sp4 (3) polar graph
27 18 18 315 −324 N U (4, 2)
+ 41 20 9 10 2.70220 −3.70220 Paley(41); 2-graph−∗
78! 45 12 3 3 320 −324 complete enumeration by Coolsaet, Degraer & Spence
[95]; U4 (2) polar graph
32 22 24 224 −420 N O5+ (3)
10
! 45 16 8 4 69 −235 2
28 15 21 135 −79
+ 45 22 10 11 2.85422 −3.85422 Mathon [241]; 2-graph−∗
! 49 12 5 2 512 −236 72
36 25 30 136 −612 OA(7,6)
− 49 16 3 6 232 −516 Bussemaker-Haemers-Mathon-Wilbrink [70]
32 21 20 416 −332
+ 49 18 7 6 418 −330 OA(7,3); Pasechnik(7)
30 17 20 230 −518 OA(7,5)
+ 49 24 11 12 324 −424 Paley(49); OA(7,4); 2-graph−∗
! 50 7 0 1 228 −321 Hoffman-Singleton graph [196]; U3 (52 ).2/Sym(7)
42 35 36 221 −328
− 50 21 4 12 142 −97 Absolute bound
28 18 12 87 −242 Absolute bound
+ 50 21 8 9 325 −424 switch skewhad2 + ∗; 2-graph
28 15 16 324 −425 S(2,4,25); 2-graph
+ 53 26 12 13 3.14026 −4.14026 Paley(53); 2-graph−∗
11
710 −244

! 55 18 9 4 2
36 21 28 144 −810
! 56 10 0 2 235 −420 Sims-Gewirtz graph [152, 153, 52];
L3 (4).22 /Alt(6).22
20 35
45 36 36 3 −3 Witt: intersection-2 graph of a 2-(21,6,4) design with
block intersections 0, 2
48
− 57 14 1 4 238 −518 Wilbrink-Brouwer [306]
42 31 30 418 −338
+ 57 24 11 9 518 −338 S(2,3,19)
32 16 20 238 −618
− 57 28 13 14 3.27528 −4.27528 Conf
+ 61 30 14 15 3.40530 −4.40530 Paley(61); 2-graph−∗
+ 63 30 13 15 335 −527 intersection-8 graph of a 2-(36,16,12) design with
block intersections 6, 8; O7 (2) polar graph; Sp6 (2)
polar graph; 2-graph−∗
32 16 16 427 −435 S(2,4,28); intersection-6 graph of a 2-(28,12,11) de-
sign with block intersections 4, 6; N U (3, 3); 2-
graph−∗
! 64 14 6 2 614 −249 82 ; from a partial spread of 3-spaces: projective bi-
nary [14,6] code with weights 4, 8
49 36 42 149 −714 OA(8,7)
167! 64 18 2 6 245 −618 complete enumeration by Haemers & Spence [184];
GQ(3, 5); from a hyperoval: projective 4-ary [6,3]
code with weights 4, 6
45 32 30 518 −345
+ 64 21 8 6 521 −342 OA(8,3); Bilin2×3 (2); from a Baer subplane: projec-
tive 4-ary [7,3] code with weights 4, 6; from a partial
spread of 3-spaces: projective binary [21,6] code with
weights 8, 12
42 26 30 242 −621 OA(8,6)
+ 64 27 10 12 336 −527 from a unital: projective 4-ary [9,3] code with weights
6, 8; V O6− (2) affine polar graph; RSHCD − ; 2-graph
36 20 20 427 −436 2-graph
+ 64 28 12 12 428 −435 OA(8,4); from a partial spread of 3-spaces: projective
binary [28,6] code with weights 12, 16; RSHCD + ; 2-
graph
35 18 20 335 −528 OA(8,5); Goethals-Seidel(2,7); V O6+ (2) affine polar
graph; 2-graph
− 64 30 18 10 108 −255 Absolute bound
? 65 32 15 16 3.53132 −4.53132 2-graph−∗?
12
811 −254

! 66 20 10 4 2
45 28 36 154 −911
? 69 20 7 5 523 −345
48 32 36 245 −623 S(2,6,46) does not exist
− 69 34 16 17 3.65334 −4.65334 Conf
+ 70 27 12 9 620 −349 S(2,3,21)
42 23 28 249 −720
+ 73 36 17 18 3.77236 −4.77236 Paley(73); 2-graph−∗
? 75 32 10 16 256 −818 2-graph−∗?
42 25 21 718 −356 2-graph−∗?
− 76 21 2 7 256 −719 Haemers [176]
54 39 36 619 −356
? 76 30 8 14 257 −818 2-graph?
45 28 24 718 −357 2-graph?
? 76 35 18 14 719 −356 2-graph?
40 18 24 256 −819 2-graph?
! 77 16 0 4 255 −621 S(3,6,22); M22 .2/24 : Sym(6); unique by Brouwer [42];
subconstituent of Higman-Sims graph; intersection-6
graph of a 2-(56,16,6) design with block intersections
4, 6
60 47 45 521 −355 Witt 3-(22,6,1): intersection-2 graph of a 2-(22,6,5)
design with block intersections 0, 2
− 77 38 18 19 3.88738 −4.88738 Conf
13
912 −265

! 78 22 11 4 2
55 36 45 165 −1012
16 64 2
! 81 16 7 2 7 −2 9 ; from a partial spread: projective ternary [8,4]
code with weights 3, 6
64 16
64 49 56 1 −8 OA(9,8)
! 81 20 1 6 260 −720 unique by Brouwer & Haemers [51]; V O4− (3) affine
polar graph; projective ternary [10,4] code with
weights 6, 9
60 45 42 620 −360
+ 81 24 9 6 624 −356 OA(9,3); V N O4+ (3) affine polar graph; from a partial
spread: projective ternary [12,4] code with weights 6,
9
8.9. TABLE 129
56 37 42 256 −724 OA(9,7)
+ 81 30 9 12 350 −630 V N O4− (3) affine polar graph; Hamada-Helleseth
[188]: projective ternary [15,4] code with weights 9,
12
50 31 30 530 −450
+ 81 32 13 12 532 −448 OA(9,4); Bilin2×2 (3); V O4+ (3) affine polar graph;
from a partial spread: projective ternary [16,4] code
with weights 9, 12
48 27 30 348 −632 OA(9,6)
− 81 40 13 26 172 −148 Absolute bound
+ 81 40 19 20 440 −540 Paley(81); OA(9,5); projective ternary [20,4] code
with weights 12, 15; 2-graph−∗
+ 82 36 15 16 441 −540 2-graph
45 24 25 440 −541 S(2,5,41); 2-graph
? 85 14 3 2 434 −350
70 57 60 250 −534
+ 85 20 3 5 350 −534 O5 (4) polar graph; Sp4 (4) polar graph
64 48 48 434 −450
? 85 30 11 10 534 −450
54 33 36 350 −634 S(2,6,51)?
? 85 42 20 21 4.11042 −5.11042 2-graph−∗?
? 88 27 6 9 355 −632
60 41 40 532 −455
+ 89 44 21 22 4.21744 −5.21744 Paley(89); 2-graph−∗
14
! 91 24 12 4 1013 −277 2
66 45 55 177 −1113
− 93 46 22 23 4.32246 −5.32246 Conf
? 95 40 12 20 275 −1019 2-graph−∗?
54 33 27 919 −375 2-graph−∗?
+ 96 19 2 4 357 −538 Haemers(4)
76 60 60 438 −457
+ 96 20 4 4 445 −450 GQ(5, 3)
75 58 60 350 −545
? 96 35 10 14 363 −732
60 38 36 632 −463
− 96 38 10 18 276 −1019 Degraer
57 36 30 919 −376
? 96 45 24 18 920 −375 2-graph?
50 22 30 275 −1020 2-graph?
+ 97 48 23 24 4.42448 −5.42448 Paley(97); 2-graph−∗
? 99 14 1 2 354 −444
84 71 72 344 −454
? 99 42 21 15 921 −377
56 28 36 277 −1021
+ 99 48 22 24 454 −644 2-graph−∗
50 25 25 544 −554 S(2,5,45); 2-graph−∗
! 100 18 8 2 818 −281 102
81 64 72 181 −918
! 100 22 0 6 277 −822 Higman-Sims graph [192]; HS01.2/M22 .2; unique by
2
Gewirtz [152]; q22 =0
22 77 1
77 60 56 7 −3 q11 =0
+ 100 27 10 6 727 −372 OA(10,3)
72 50 56 272 −827 OA(10,8)?
? 100 33 8 12 366 −733
66 44 42 633 −466
+ 100 33 14 9 824 −375 S(2,3,25)
66 41 48 275 −924
− 100 33 18 7 1311 −288 Absolute bound
66 39 52 188 −1411 Absolute bound
+ 100 36 14 12 636 −463 Hall-Janko graph; J2 .2/U3 (3).2; subconstituent of
G2 (4) graph; OA(10,4)
63 38 42 363 −736 OA(10,7)?
+ 100 44 18 20 455 −644 Jørgensen-Klin graph [209]; RSHCD − ; 2-graph
∃ k λ µ rf
v sg comments
55 30 30 544 −555 2-graph
+ 100 45 20 20 545 −554 OA(10,5)?; RSHCD + ; 2-graph
54 28 30 454 −645 OA(10,6)?; 2-graph
Comments
Comment Explanation
1 2
q11 = 0, q22 =0 zero Krein parameter, see §10.4
2
m Hamming graph H(2, m), i.e., lattice graph L2 (m),
i.e., grid graph m × m, i.e., Km × Km , see §11.3.1,
§1.4.5
m
2 Johnson graph J(m, 2), i.e., triangular graph T (m),
see §11.3.2, §1.4.5
OA(n, t) (t ≥ 3) block graph of an orthogonal array OA(n, t) (t − 2
mutually orthogonal Latin squares of order n)
S(2, k, v) block graph of a Steiner system S(2, k, v), i.e., of a
2-(v, k, 1) design
Goethals-Seidel(k, r) graph constructed from a Steiner system S(2, k, v)
(with r = (v − 1)/(k − 1)) and a Hadamard matrix
of order r + 1 as in [160]
2-graph graph in the switching class of a regular 2-graph, see
§9.2
2-graph−∗ descendant of a regular 2-graph, see §9.2
RSHCD± Regular 2-graph derived from a regular symmetric
Hadamard matrix with constant diagonal (cf. §9.5,
[56], [160])
Taylor 2-graph for U3 (q) graph derived from Taylor’s regular 2-graph (cf. [56],
[294], [295])
Paley(q) Paley graph on Fq , see §9.4, §12.6
vanLint-Schrijver(u) graph constructed by the cyclotomic construction of
[228], taking the union of u classes
Bilin2×d (q) graph on the 2 × d matrices over Fq , adjacent when
their difference has rank 1
GQ(s, t) collinearity graph of a generalized quadrangle with pa-
rameters GQ(s, t), see §8.6.3
ε
O2d (q), O2d+1 (q) isotropic points on a nondegenerate quadric in the pro-
jective space P G(2d − 1, q) or P G(2d, q), joined when
the connecting line is totally singular
Sp2d (q) points of P G(2d−1, q) provided with a nondegenerate
symplectic form, joined when the connecting line is
totally isotropic
Ud (q) isotropic points of P G(d − 1, q 2 ) provided with a non-
degenerate Hermitean form, joined when the connect-
ing line is totally isotropic
continued...
8.10. EXERCISES 131
Comment Explanation
ε
N O2d (2) nonisotropic points of P G(2d − 1, 2) provided with a
nondegenerate quadratic form, joined when they are
orthogonal, i.e., when the connecting line is a tangent
ε
N O2d (3) one class of nonisotropic points of P G(2d − 1, 3) pro-
vided with a nondegenerate quadratic form, joined
when they are orthogonal, i.e., when the connecting
line is elliptic
ε
N O2d+1 (q) one class of nondegenerate hyperplanes of P G(2d, q)
provided with a nondegenerate quadratic form, joined
when their intersection is degenerate
ε⊥
N O2d+1 (5) one class of nonisotropic points of P G(2d, 5) provided
with a nondegenerate quadratic form, joined when
they are orthogonal
N Un (q) nonisotropic points of P G(n − 1, q) provided with a
nondegenerate Hermitean form, joined when the con-
necting line is a tangent
ε
V O2d (q) vectors of a 2d-dimensional vector space over Fq pro-
vided with a nondegenerate quadratic form Q, where
two vectors u and v are joined when Q(v − u) = 0
ε
V N O2d (q) (q odd) vectors of a 2d-dimsensional vector space over Fq pro-
vided with a nondegenerate quadratic form Q, where
two vectors u and v are joined when Q(v − u) is a
nonzero square
8.10 Exercises
Exercise 1 ([160]) Consider the graph on the set of flags (incident point-line
pairs) of the projective plane P G(2, 4) where (p, L) and (q, M ) are adjacent
when p 6= q and L 6= M and either p ∈ M or q ∈ L. Show that this graph is
strongly regular with parameters (v, k, λ, µ) = (105, 32, 4, 12).
Exercise 2 ([21]) Consider the graph on the cosets of the perfect ternary Golay
code (an [11,6,5] code over F3 ), where two cosets are adjacent when they differ
by a vector of weight 1. Show that this graph is strongly regular with parameters
(v, k, λ, µ) = (243, 22, 1, 2). It is known as the Berlekamp-van Lint-Seidel graph.
Exercise 3 For a strongly regular graph Γ and a vertex x of Γ, let ∆ be the
subgraph of Γ induced on the set of vertices different from x and nonadjacent
to x. If Γ has no triangles and spectrum k 1 , rf , sg , then show that ∆ has
spectrum (k − µ)1 , rf −k , sg−k , (−µ)k−1 . Conclude if Γ is primitive that f ≥ k
and g ≥ k, and that if f = k or g = k then ∆ is itself complete or strongly
regular. Determine all strongly regular graphs with λ = 0 and f = k.
Exercise 4 ([32]) Show that having a constant k almost follows from having
constant λ, µ. More precisely: Consider a graph Γ with the property that any
two adjacent (non-adjacent) vertices have λ (resp. µ) common neighbors. Show
that if Γ is not regular, then either µ = 0 and Γ is a disjoint union of (λ + 2)-
cliques, or µ = 1, and Γ is obtained from a disjoint union of (λ + 1)-cliques by
adding a new vertex, adjacent to all old vertices.
Exercise 5 Prove Theorem 8.6.2.

Exercise 6 A spread in a generalized quadrangle is a subset S of the lines
such that every point is on exactly one line of S. Prove that a GQ(q 2 , q) has no
spread. Hint: A spread is a coclique in the line graph.
Exercise 7 ([209]) Show that the strongly regular graph with parameters
(v, k, λ, µ) = (100, 45, 20, 20) obtained from the Hall-Janko graph in §8.1.12
can be switched into a strongly regular graph with parameters (100,55,30,30).
Exercise 8 There exist strongly regular graphs in F43 , invariant for translation
and dilatation, with parameters (v, k, λ, µ) = (81, 20, 1, 6) and (81, 30, 9, 12).
Determine the corresponding ternary codes and their weight enumerators.
Exercise 9 With C and D as in §8.7, show that C∪D induces a distance-regular
graph of diameter three with intersection array {10, 9, 4; 1, 6, 10}.
Exercise 10 With Γ as in §8.7, show that χ(Γ) ≥ 6 also follows from Corol-
lary 4.2.4 applied to the induced subgraph of Γ, obtained by deleting all vertices
of one color class.
Exercise 11 Under what conditions is the Hamming code cyclic? Negacyclic?
Constacyclic?
Exercise 12 A cap in a projective space is a collection of points, no three on a
line. Show that a [n, n − m, 4] code over Fq exists if and only if there is a cap of
size n in P G(m − 1, q). Construct for m > 0 a [2m−1 , 2m−1 − m, 4] binary code.
Exercise 13 Given a two-weight code over Fq of word length n, dimension m
and weights w1 and w2 . Express the parameters v, k, λ, µ, r, s, f , g of the
corresponding strongly regular graph in terms of q, n, k, w1 and w2 .
Chapter 9
Regular two-graphs
9.1 Strong graphs

Let us call a graph (possibly improper) strongly regular when it is strongly reg-
ular or complete or edgeless. Above (Theorem 8.1.2) we saw that a graph Γ is
(possibly improper) strongly regular if and only if its adjacency matrix A satis-
fies A2 ∈ hA, I, Ji, where h...i denotes the R-span. In particular, this condition
implies that Γ is regular, so that AJ = JA.
Consider the Seidel matrix S = J − I − 2A (see §1.8.2). We have hA, I, Ji =
hS, I, Ji. If A2 ∈ hA, I, Ji then also S 2 ∈ hS, I, Ji, but the converse does not
hold. For example, consider the path P3 of length 2. We have S 2 = S + 2I, but
A only satisfies the cubic equation A3 = 2A.
We call a graph strong whenever its Seidel matrix S satisfies S 2 ∈ hS, I, Ji.
Thus a (possibly improper) strongly regular graph is strong, and conversely
a regular strong graph is (possibly improper) strongly regular. As we saw, a
strong graph need not be regular. Another example is given by C5 + K1 , where
the Seidel matrix satisfies S 2 = 5I. But the following properties are satisfied
(recall that an eigenvalue is called restricted if it has an eigenvector orthogonal
to the all-ones vector 1):
Proposition 9.1.1 For a graph Γ with v vertices and Seidel matrix S the fol-
lowing holds:
(i) Γ is strong if and only if S has at most two restricted eigenvalues. In this
case (S − ρ1 I)(S − ρ2 I) = (v − 1 + ρ1 ρ2 )J, where ρ1 and ρ2 are restricted
eigenvalues of S.
(ii) Γ is strong and regular if and only if Γ is (possibly improper) strongly

regular. In this case the eigenvalue ρ0 of S for 1 satisfies (ρ0 − ρ1 )(ρ0 −
ρ2 ) = v(v − 1 + ρ1 ρ2 ).
(iii) If Γ is strong with restricted eigenvalues ρ1 and ρ2 , and v − 1 + ρ1 ρ2 6= 0,

then Γ is regular, and hence (possibly improper) strongly regular.
(iv) S has a single restricted eigenvalue if and only if S = ±(J − I), that is, if
and only if Γ is complete or edgeless.
133
134 CHAPTER 9. REGULAR TWO-GRAPHS
Proof. (i) If Γ is strong then S 2 + αS + βI = γJ for some constants α, β and

γ. If ρ is a restricted eigenvalue of S with eigenvector v orthogonal to 1, then
(ρ2 + αρ + β)v = γJv = 0, so ρ2 + αρ + β = 0. Therefore S has at most two re-
stricted eigenvalues. Conversely, if S has just two restricted eigenvalues ρ1 and
ρ2 , then (S −ρ1 I)(S −ρ2 I) ∈ hJ i, so Γ is strong. And if (S −ρ1 I)(S −ρ2 I) = γJ,
then the diagonal entries show that γ = v − 1 + ρ1 ρ2 .
(ii) We know that (possibly improper) strongly regular implies strong and reg-
ular. Suppose Γ is strong and regular, then S 2 ∈ hS, I, Ji and SJ ∈ hJ i, this
implies that the adjacency matrix A = (J −S −I)/2 of Γ satisfies A2 ∈ hA, I, Ji,
so Γ is (possibly improper) strongly regular by Theorem 8.1.2.
(iii) If Γ is not regular, then J is not a polynomial in S, so v − 1 + ρ1 ρ2 = 0
follows from part (i).
We see that v − 1 + ρ1 ρ2 = 0 if and only if S has exactly two distinct eigenvalues
ρ1 and ρ2 . Recall that two graphs Γ and Γ e are switching equivalent (see Sec-
e
tion 1.8.2) if their Seidel matrices S and S are similar by some diagonal matrix
D = diag(±1, . . . , ±1) (i.e. Se = DSD). So switching equivalent graphs have
the same Seidel spectrum, and therefore the property of being strong with two
Seidel eigenvalues is invariant under Seidel switching.
Suppose Γ is a strong graph on v vertices with two Seidel eigenvalues ρ1
and ρ2 (so v − 1 + ρ1 ρ2 = 0). Clearly, Γ is regular of degree k if and only if
its Seidel matrix has constant row sum v − 1 − 2k. Therefore v − 1 − 2k = ρ0
is an eigenvalue of S, so either ρ0 = ρ1 , or ρ0 = ρ2 . Switching in Γ produces
another strong graph, which may or may not be regular. If it is regular, then it
is regular of degree either (v − 1 − ρ1 )/2 or (v − 1 − ρ2 )/2.
Examples (i) If Γ is P3 , then the Seidel eigenvalues are −1 and 2, so a regular
graph that is switching equivalent must have degree either 3/2 or 0. The former
is impossible, but the latter happens: P3 is switching √ equivalent to 3K1 .
(ii) If Γ is C5 + K1 , then the eigenvalues are ± 5, and so can never be equal
to the row sum. So this graph cannot be switched into a regular one.
(iii) If Γ is the 4×4 grid (the lattice graph L2 (4)), then v = 16 and ρ0 = ρ1 =
3, ρ2 = −5. So Γ is strong with two eigenvalues. Switching in Γ with respect to
a coclique of size 4 gives again a regular graph with the same parameters as Γ,
but which is not isomorphic to Γ. This is the Shrikhande graph (see Section 8.2).
Switching with respect to the union of two parallel lines in the grid (that is, two
disjoint 4-cliques in Γ) gives a regular graph of degree 10, the Clebsch graph
(see Section 8.2).
Strong graphs were introduced by Seidel [274].
9.2 Two-graphs
A two-graph Ω = (V, ∆) consists of a finite set V , together with a collection
∆ of unordered triples from V , such that every 4-subset of V contains an even
number of triples from ∆. The triples from ∆ are called coherent.
From a graph Γ = (V, E), one can construct a two-graph Ω = (V, ∆) by
defining a triple from V to be coherent if the three vertices induce a subgraph
in Γ with an odd number of edges. It is easily checked that out of the four triples
in any graph on four vertices, 0, 2, or 4 are coherent. So Ω is a two-graph. We
call Ω the two-graph associated to Γ.
9.3. REGULAR TWO-GRAPHS 135
Observe that Seidel switching does not change the parity of the number of
edges in any 3-vertex subgraph of Γ. Therefore switching equivalent graphs have
the same associated two-graph. Conversely, from any two-graph Ω = (V, ∆)
one can construct a graph Γ as follows. Take ω ∈ V . Define two vertices
x, y ∈ V \ {ω} to be adjacent in Γ if {ω, x, y} ∈ ∆, and define ω to be an isolated
vertex of Γ. We claim that every triple {x, y, z} ∈ ∆ has an odd number of edges
in Γ, which makes Ω the two-graph associated to Γ. If ω ∈ {x, y, z} this is clear.
If ω 6∈ {x, y, z}, the 4-subgraph condition implies that {x, y, z} ∈ ∆ whenever
from the triples {ω, y, z}, {ω, x, y}, {ω, x, z} just one, or all three are coherent.
Hence {x, y, x} has one or three edges in Γ. Thus we have established a one-to-
one correspondence between two-graphs and switching classes of graphs.
Small two-graphs were enumerated in [71]. The number of nonisomorphic
two-graphs on n vertices for small n is
n 0 1 2 3 4 5 6 7 8 9 10
# 1 1 1 2 3 7 16 54 243 2038 33120
There is an explicit formula for arbitrary n. See, e.g., [237].

For the graph Γ with an isolated vertex ω, obtained from Ω as indicated
above, the graph Γ \ ω plays an important role. It is called the descendant of Ω
with respect to ω, and will be denoted by Γω .
Since switching equivalent graphs have the same Seidel spectrum, we can
define the eigenvalues of a two-graph to be the Seidel eigenvalues of any graph
in the corresponding switching class.
Seidel & Tsaranov [278] classified the two-graphs with smallest Seidel eigen-
value not less than −3:
Theorem 9.2.1 (i) A graph Γ with smallest Seidel eigenvalue larger than −3
is switching equivalent to the void graph on n vertices, to the one-edge graph
on n vertices, or to one of the following 2 + 3 + 5 graphs on 5, 6, 7 vertices,
respectively:
• • • • • • • • • •
• •• • • • • •• • • • • • • • • • • • • • • • •• •
• • • • • • •• • • • • ••• ••• ••• ••• • •
(ii) A graph Γ with smallest Seidel eigenvalue not less than −3 is switching
equivalent to a subgraph of mK2 or of T (8), the complement of the line graph
of K8 .
9.3 Regular two-graphs

A two-graph (V, ∆) is called regular (of degree a) if every unordered pair from V
is contained in exactly a triples from ∆. Suppose Ω = (V, ∆) is a two-graph, and
let ∇ be the set of non-coherent triples, then it easily follows that Ω = (V, ∇)
is also a two-graph, called the complement of Ω. Moreover, Ω is regular of
degree a if and only if the complement Ω is regular of degree a = v − 2 − a.
The following result relates regular two-graphs with strong graphs and strongly
regular graphs.
Theorem 9.3.1 For a graph Γ with v vertices, its associated two-graph Ω, and
any descendant Γω of Ω the following are equivalent.
(i) Γ is strong with two Seidel eigenvalues ρ1 and ρ2 .
(ii) Ω is regular of degree a.
(iii) Γω is (possibly improper) strongly regular with parameters (v − 1, k, λ, µ)

with µ = k/2.
The parameters are related by v = 1 − ρ1 ρ2 , a = k = 2µ = −(ρ1 + 1)(ρ2 + 1)/2,

and λ = (3k − v)/2 = 1 − (ρ1 + 3)(ρ2 + 3)/4. The restricted Seidel eigenvalues
of Γω are ρ1 and ρ2 , and ρ1 + ρ2 = v − 2a − 2 = a − a.
Proof. (ii) ⇒ (iii): Let x be a vertex of Γω . The number of coherent triples

containing ω and x equals the number of edges in Γω containing x, so Γω is
regular of degree a. For two vertices x and y in Γω , let p(x, y) denote the number
of vertices z (z 6= x, y) adjacent to x but not to y. If x and y are distinct non-
adjacent, then p(x, y) + p(y, x) = a, and the number µ of common neighbors of
x and y equals k −p(x, y) = k −p(y, x). Therefore µ = k/2 = a/2 is independent
of x and y. Similarly, if x and y are adjacent, then p(x, y) + p(y, x) = a (the
degree of the complement), and the number λ of common neighbors of x and y
equals k − 1 − p(x, y) = k − 1 − p(y, x), which implies λ = (3k − v)/2, which is
independent of x and y.
(iii) ⇒ (ii): If Γω is strongly regular and k = 2µ, then Theorem 8.1.3 gives
λ = (3k − v)/2. With the relations above this shows that Ω is regular of
degree k.
(i) ⇒ (iii): Switch in Γ with respect to the neigbors of ω, then ω becomes
isolated, and Γ \ ω = Γω . If Sω is the Seidel matrix of Γω , then

0 1⊤
S=
1 Sω
is the Seidel matrix of Γ. We know (S − ρ1 I)(S − ρ2 I) = 0. This gives (Sω −

ρ1 I)(Sω − ρ2 I) = −J, Therefore Γω is strongly regular with restricted Seidel
eigenvalues ρ1 and ρ2 and v − 1 = −ρ1 ρ2 vertices. From S = J − 2A − I we get
the adjacency eigenvalues r = −(ρ1 + 1)/2 and s = −(ρ2 + 1)/2 of Γω . Now the
parameters of Γω follow from Theorem 8.1.3.
(iii) ⇒ (i): Suppose Γω is strongly regular with k = 2µ and Seidel matrix Sω .
Then it follows readily that Sω 1 = (ρ1 + ρ2 )1 and (Sω − ρ1 I)(Sω − ρ2 I) = −J.
This implies that S satisfies (S − ρ1 I)(S − ρ2 I) = 0.
Small regular two-graphs have been classified. The table below gives the
numbers of nonisomorphic nontrivial regular two-graphs with ρ1 = −3 or ρ1 =
−5 or v ≤ 50.
v 6
√ 10 14
√ 16 18
√ 26 28 30
√ 36
ρ1 , ρ2 ± 5 ±3 ± 13 −3, 5 ± 17 ±5 −3, 9 ± 29 −5, 7
# 1 1 1 1 1 4 1 6 227
v 38
√ 42
√ 46
√ 50 76 96 126 176 276
ρ1 , ρ2 ± 37 ± 41 ± 45 ±7 −5, 15 −5, 19 −5, 25 −5, 35 −5, 55
# ≥ 191 ≥ 18 ≥ 97 ≥ 54 ? ? 1 1 1
9.3. REGULAR TWO-GRAPHS 137
9.3.1 Related strongly regular graphs

Given the parameters of a regular two-graph Ω, we find three parameter sets for
strongly regular graphs that may be related, namely that of the descendants,
and the two possible parameter sets for regular graphs in the switching class of
Ω. The parameters are given by:
Proposition 9.3.2 (i) Let Γ be strongly regular with parameters (v, k, λ, µ).
The associated two-graph Ω is regular if and only if v = 2(2k − λ − µ). If this
is the case, then it has degree a = 2(k − µ), and Γω is strongly regular with
parameters (v − 1, 2(k − µ), k + λ − 2µ, k − µ).
(ii) Conversely, if Γ is regular of valency k, and the associated two-graph Ω
is regular of degree a, then Γ is strongly regular with parameters λ = k−(v−a)/2
and µ = k − a/2, and k satisfies the quadratic 2k 2 − (v + 2a)k + (v − 1)a = 0.
Proof. (i) By definition, Ω is regular of degree a if and only if a = λ + (v −

2k + λ) = 2(k − µ). The parameters follow immediately.
(ii) The quadratic expresses that k − 21 v ∈ {r, s}.
In the case of the regular two-graph on 6 vertices, the descendants are pentagons,
and there are no regular graphs in the switching class.
In the case of the regular two-graph on 10 vertices, the descendants are
grid graphs 3 × 3. The switching class contains both the Petersen graph and
its complement. Therefore Ω is isomorphic to its complement (and so are the
descendants).
In the case of the regular two-graph on 16 vertices, the descendants are
isomorphic to the triangular graph T (6) (with parameters (15,8,4,4) and spec-
trum 81 25 (−2)9 ). The switching class contains the grid graph 4 × 4 and
the Shrikhande graph (both with parameters (16,6,2,2) and spectrum 61 26
(−2)9 ), and the Clebsch graph (with parameters (16,10,6,6) and spectrum 101
25 (−2)10 ).
It remains to specify what switching sets are needed to switch between two
strongly regular graphs associated to the same regular two-graph.
Proposition 9.3.3 Let Γ be strongly regular with parameters (v, k, λ, µ), asso-
ciated with a regular two-graph.
(i) The graph Γ is switched into a strongly regular graph with the same pa-
rameters if and only if every vertex outside the switching set S is adjacent
to half of the vertices of S.
(ii) The graph Γ is switched into a strongly regular graph with parameters
(v, k + c, λ + c, µ + c) where c = 21 v − 2µ if and only if the switching set S
has size 12 v and is regular of valency k − µ.
For example, in order to switch the 4 × 4 grid graph into the Shrikhande
graph, we can switch with respect to a 4-coclique. And in order to switch the
4 × 4 grid graph into the Clebsch graph, we need a split into two halves that
are regular with valency 4, and the union of two disjoint K4 ’s works.
Regular two-graphs were introduced by Graham Higman and further inves-
tigated by Taylor [293].
9.3.2 The regular two-graph on 276 points

If N is the point-block incidence matrix of the unique Steiner system S(4, 7, 23),
then N N ⊤ = 56I + 21J, N J = 77J, JN = 7J. Since any two blocks in this
Steiner system meet in 1 or 3 points, we have N ⊤ N = 7I + A + 3(J − I − A)
where A describes the relation of meeting in 1 point. As we already saw in
§8.1.10, A is the adjacency matrix of a strongly regular graph—in this case
one with parameters (v, k, λ, µ) = (253, 112, 36, 60) and spectrum 1121 2230
(−26)22 . The Seidel matrix S = J − I − 2A has 1 230
spectrum 28 (−5) 51 and
22
J −I J − 2N
satisfies (S − 51I)(S + 5I) = −3J. Now S ′ = satisfies
J − 2N ⊤ S
(S ′ − 55I)(S ′ + 5I) = 0 and hence is the Seidel matrix of a regular two-graph
on 276 vertices. This two-graph is unique (Goethals & Seidel [161]). Its group
of automorphisms is Co3 , acting 2-transitively.
9.3.3 Coherent subsets

A clique, or coherent subset in a two-graph Ω = (V, ∆) is a subset C of V
such that all triples in C are coherent. If x ∈/ C, then x determines a partition
{Cx , Cx′ } of C into two possibly empty parts such that a triple xyz with y, z ∈ C
is coherent precisely when y and z belong to the same part of the partition.
Proposition 9.3.4 (Taylor [295]) Let C be a nonempty coherent subset of the

regular two-graph Ω with eigenvalues ρ1 , ρ2 , where ρ2 < 0. Then
/ C we have |Cx | = |Cx′ |,
(i) |C| ≤ 1 − ρ2 , with equality iff for each x ∈
and
(ii) |C| ≤ m(ρ2 ).
Proof. (i) Let c = |C|.P Counting incoherent P triples 2that 1meet C in two points,
we find 12 c(c − 1)a = x∈C / |C x |.|C ′
x | ≤ x∈C
/ (c/2) = 4 c 2
(v − c). It follows
that c2 − (v − 2a)c − 2a ≤ 0. But the two roots of x2 − (v − 2a)x − 2a = 0 are
1 − ρ1 and 1 − ρ2 , hence 1 − ρ1 ≤ c ≤ 1 − ρ2 .
(ii) This follows by making a system of equiangular lines in Rm as in §9.6.1
corresponding to the complement of Ω. We can choose unit vectors for the
points in C such that their images form a simplex (any two have the same inner
product) and hence |C| is bounded by the dimension m = v − m(ρ1 ) = m(ρ2 ).
9.3.4 Completely regular two-graphs

In a regular two-graph each pair is in a2 = a coherent triples, that is, in a2
3-cliques, and each coherent triple is in a3 4-cliques, where a3 is the number of
common neighbours of two adjacent vertices in any strongly regular graph Γω ,
so that a3 = − 41 (ρ1 + 3)(ρ2 + 3) + 1 by Theorem 9.3.1.
Let a t-regular two-graph be a regular two-graph in which every i-clique is
contained in a nonzero constant number ai of (i + 1)-cliques, for 2 ≤ i ≤ t.
By Proposition 9.3.4 we must have t ≤ −ρ2 . A completely regular two-graph
is a t-regular two-graph with t = −ρ2 . For example, the regular two-graph
on 276 points (§9.3.2) is completely regular. Neumaier [252] introduced this
concept and gave parameter restrictions strong enough to leave only a finite list
of feasible parameters. There are five examples, and two open cases.
9.4. CONFERENCE MATRICES 139
# ρ1 ρ2 v a2 a3 a4 a5 a6 a7 existence
1 3 −3 10 4 1 unique [273]
2 5 −3 16 6 1 unique [273]
3 9 −3 28 10 1 unique [273]
4 7 −5 36 16 6 2 1 unique (BH)
5 19 −5 96 40 12 2 1 none (NP)
6 25 −5 126 52 15 2 1 none [252]
7 55 −5 276 112 30 2 1 unique [161]
8 21 −7 148 66 25 8 3 2 1 none [252]
9 41 −7 288 126 45 12 3 2 1 none [28]
10 161 −7 1128 486 165 36 3 2 1 ?
11 71 −9 640 288 112 36 10 4 3 none (BH)
12 351 −9 3160 1408 532 156 30 4 3 ?
13 253 −11 2784 1270 513 176 49 12 5 none [252]
Here (BH) refers to an unpublished manuscript by Blokhuis and Haemers, while
(NP) is the combination of Neumaier [252] who showed that a derived graph
on 95 vertices must be locally GQ(3, 3), and Pasechnik [254] who classified such
graphs and found none on 95 vertices.
9.4 Conference matrices

The Seidel matrix of C5 + K1 is an example of a so called conference matrix. An
n×n matrix S is a conference matrix if all diagonal entries are 0, the off-diagonal
entries are ±1, and SS ⊤ = (n − 1)I.
Multiplying a row or column by −1 (switching) does not affect the conference
matrix property. It was shown in [128] that any conference matrix can be
switched into a form where it is either symmetric or skew symmetric:
Lemma 9.4.1 Let S be a conference matrix of order n with n > 2. Then n

is even and one can find diagonal matrices D and E with diagonal entries ±1
such that (DSE)⊤ = DSE if and only if n ≡ 2 (mod 4). One can find such D
and E with (DSE)⊤ = −DSE if and only if n ≡ 0 (mod 4).
Proof. Switch rows and columns so as to make all non-diagonal entries of the
first row and column equal to 1. The second row now has n/2 entries 1 and
equally many entries −1 (since it has inner product zero with the first row).
So, n is even, say n = 2m + 2. Let there be a, b, c, d entries 1, −1, 1, −1 in the
third row below the entries 1, 1, −1, −1 of the second row, respectively. We may
assume (by switching the first column and all rows except the first if required)
that S23 = 1. If S32 = 1 then a + b = m − 1, c + d = m, a + c + 1 = m,
a − b − c + d + 1 = 0 imply a + 1 = b = c = d = 21 m so that m is even. If
S32 = −1 then a + b = m − 1, c + d = m, a + c = m, a − b − c + d + 1 = 0
imply a = b = c − 1 = d = 12 (m − 1) so that m is odd. This proves that after
switching the first row and column to 1, the matrix S has become symmetric in
case n ≡ 2 (mod 4), while after switching the first row to 1 and the first column
to −1, the matrix S has become skew symmetric in case n ≡ 0 (mod 4).
Thus, if n ≡ 2 (mod 4), S gives rise to a strong graph with two eigenvalues and
its associated two-graph is regular of degree (n − 2)/2. The descendants are
strongly regular with parameters (n − 1, (n − 2)/2, (n − 6)/4, (n − 2)/4). We call
these graphs conference graphs. Conference graphs are characterized among the
strongly regular graphs by f = g (f and g are the multiplicities of the restricted
eigenvalues), and are the only cases in which non-integral eigenvalues can occur.
The following condition is due to Belevitch [20].
Theorem 9.4.2 If n is the order of a symmetric conference matrix, then n − 1
is the sum of two integral squares.
Proof. CC ⊤ = (n − 1)I implies that I and (n − 1)I are rationally congruent
(two matrices A and B are rationally congruent if there exist a rational matrix
R such that RAR⊤ = B). A well-known property (essentially Lagrange’s four
squares theorem) states that for every positive rational number α, the 4 × 4
matrix αI4 is rationally congruent to I4 . This implies that the n × n matrix
αIn is rationally congruent to diag(1, . . . , 1, α, . . . , α) where the number of ones
is divisible by 4. Since n ≡ 2 (mod 4), I must be rationally congruent to
diag(1, . . . , 1, n − 1, n − 1). This implies that n − 1 is the sum of two squares.
Note that this theorem also gives a necessary condition for the existence of
conference graphs. For example, 21 is not the sum of two squares, therefore
there exists no conference matrix of order 22, and no strongly regular graph
with parameters (21, 10, 4, 5).
For many values of n conference matrices are known to exist, see for example
[159]. The following construction, where n − 1 is an odd prime power, is due
to Paley [253]. Let Sω be a matrix whose rows and columns are indexed by the
elements of a finite field Fq of order q, q odd. by (Sω )i,j = χ(i − j), where χ is
the quadratic residue character (that is, χ(0) = 0 and χ(x) = 1 if x is a square,
and −1 if x is not a square). It follows that S is symmetric if q ≡ 1 (mod 4),
and S is skew symmetric if q ≡ 3 (mod 4). In both cases

0 1⊤
S=
1 Sω
is a conference matrix. If n ≡ 2 (mod 4), S represents a regular two-graph and

all its descendants are isomorphic. They are the Paley graphs, that we already
encountered in §8.1.2.
9.5 Hadamard matrices

Closely related to conference matrices are Hadamard matrices. A matrix H of
order n is called a Hadamard matrix if every entry is 1 or −1, and HH ⊤ = nI.
If H is a Hadamard matrix, then so is H ⊤ . If a row or a column of a Hadamard
matrix is multiplied by −1, the matrix remains a Hadamard matrix. The core of
a Hadamard matrix H (with respect to the first row and column) is the matrix
C of order n − 1 obtained by first multiplying rows and columns of H by ±1 so
as to obtain a Hadamard matrix of which the first row and column consist of
ones only, and then deleting the first row and column. Now all entries of C are
±1, and we have CC ⊤ = C ⊤ C = nI − J, and C1 = C ⊤ 1 = −1. This implies
that the (0, 1) matrix N = 21 (C + J) satisfies N ⊤ 1 = ( 12 n − 1)1 and N N ⊤ =
1 1
4 nI + ( 4 n − 1)J, so that, for n > 2, N is the incidence matrix of a symmetric
2-(n − 1, 12 n − 1, 41 n − 1) design. Conversely, if N is the incidence matrix of a
2-design with these parameters, then 2N − J is the core of a Hadamard matrix.
9.5. HADAMARD MATRICES 141
Note that the design parameters imply that n is divisible by 4 if n > 2. The
famous Hadamard conjecture states that this condition is sufficient for existence
of a Hadamard matrix of order n. Many constructions are known (see below),
but the conjecture is still far from being solved.
A Hadamard matrix H is regular if H has constant row and column sum
(ℓ say). Now −H is a regular Hadamard √ matrix with row sum −ℓ. From
HH ⊤ = nI we get that ℓ2 = n, so ℓ = ± n, and n is a square. If H is a
regular Hadamard matrix with row sum ℓ, then N = 12 (H + J) is the incidence
matrix of a symmetric 2-(n, (n + ℓ)/2, (n + 2ℓ)/4) design. Conversely, if N is the
incidence matrix of a 2-design with these parameters (a Menon design), then
2N − J is a regular Hadamard matrix.
A Hadamard matrix H is graphical if it is symmetric with constant diagonal.
Without loss of generality we assume that the diagonal elements are 1 (otherwise
we replace H by −H). If H is a graphical Hadamard matrix of order n, then
S=H √− I is the Seidel matrix of a strong graph Γ with two Seidel eigenvalues:
−1 ± n. In other words, Γ is in the switching class of a regular two-graph. The
descendent of Γ with respect to some vertex has Seidel matrix C − I, where C
is the corresponding core of H. It is a strongly regular graph with parameters
(v, k, λ, µ) = (n − 1, 12 n − 1, 14 n − 1, 41 n − 1). From tr S = 0 it follows that also
for a graphical Hadamard √ matrix n is a square. If, in addition, H is regular
with row sum ℓ = ± n, then Γ is a strongly regular graph with parameters
(n, (n − ℓ)/2, (n − 2ℓ)/4, (n − 2ℓ)/4). And conversely, a strongly regular graph
with one of the above parameter sets gives rise to a Hadamard matrix of order
n.
There is an extensive literature on Hadamard matrices. See, e.g., [271, 272,
100].
Constructions
There is a straightforward construction of Hadamard matrices from conference
matrices. If S is a skew symmetric conference matrix, then H = S + I is a
Hadamard matrix, and if S is a symmetric conference matrix, then

S+I S−I
H=
S − I −S − I
is a Hadamard matrix. Thus the conference matrices constructed in the previous

section give Hadamard matrices of order n = 4m if 4m − 1 is a prime power,
and if m is odd and 2m − 1 is a prime power. Some small Hadamard matrices
are:
   
1 1 1 −1 1 −1 −1 −1
 1 1 −1 1   1 −1 −1 
1 1
,   and  −1  .
1 −1  1 −1 1 1   −1 −1 1 −1 
−1 1 1 1 −1 −1 −1 1
Observe that the two Hadamard matrices of order 4 are regular and graphical.
One easily verifies that, if H1 and H2 are Hadamard matrices, then so is the
Kronecker product H1 ⊗ H2 . Moreover, if H1 and H2 are regular with row sums
ℓ1 and ℓ2 , respectively, then H1 ⊗H2 is regular with row sum ℓ1 ℓ2 . Similarly, the
Kronecker product of two graphical Hadamard matrices is graphical again. With
the small Hadamard matrices given above, we can make Hadamard matrices of
order n = 2t and regular graphical Hadamard matrices of order n = 4t with row
sum ℓ = ±2t .
Let RSHCD be the set of pairs (n, ε) such that√there exists a regular sym-
metric Hadamard matrix H with row sums ℓ = ε n and constant diagonal,
with diagonal entries 1. If (m, δ), (n, ε) ∈ RSHCD, then (mn, δε) ∈ RSHCD.
We mention some direct constructions:
(i) (4, ±1), (36, ±1), (100, ±1), (196, ±1) ∈ RSHCD.
(ii) If there exists a Hadamard matrix of order m, then (m2 , ±1) ∈ RSHCD.
(iii) If both a − 1 and a + 1 are odd prime powers, then (a2 , 1) ∈ RSHCD.
(iv) If a + 1 is a prime power and there exists a symmetric conference matrix
of order a, then (a2 , 1) ∈ RSHCD.
(v) If there is a set of t − 2 mutually orthogonal Latin squares of order 2t,
then (4t2 , 1) ∈ RSHCD.
(vi) (4t4 , ±1) ∈ RSHCD.
See [160], [56] and [271], §5.3. For the third part of (i), see [209]. For the
fourth part of (i), cf. [160], Theorem 4.5 (for k = 7) and [204]. For (ii), cf. [160],
Theorem 4.4, and [179]. For (iii), cf. [271], Corollary 5.12. For (iv), cf. [271],
Corollary 5.16. For (v), consider the corresponding Latin square graph. For
(vi), see [186].
9.6 Equiangular lines

9.6.1 Equiangular lines in Rd and two-graphs
Seidel (cf. [221, 229, 129]) studied systems of lines in Euclidean space Rd , all
passing through the origin 0, with the property that any two make the same
angle ϕ. The cases ϕ = 0 (only one line) and ϕ = π2 (at most d lines, mutually
orthogonal) being trivial, we assume 0 < ϕ < π2 . Let α = cos ϕ, so that
0 < α < 1. Choose for each line ℓi a unit vector xi on ℓi (determined up to
sign). Then x⊤ ⊤
i xi = 1 for each i, and xi xj = ± cos ϕ = ±α for i 6= j.
For the Gram matrix G of the vectors xi this means that G = I + αS, where
S is the Seidel adjacency matrix of a graph Γ. (That is, S is symmetric with
zero diagonal, and has entries −1 and 1 for adjacent and nonadjacent vertices,
respectively.) Note that changing the signs of some of the xi corresponds to
Seidel switching of Γ.
Conversely, let S be the Seidel adjacency matrix of a graph on at least two
vertices, and let θ be the smallest eigenvalue of S. (Then θ < 0 since S 6= 0 and
tr S = 0.) Now S − θI is positive semi-definite, and G = I − θ1 S is the Gram
matrix of a set of vectors in Rd , where d = rk (S − θI) = n − m(θ) where n is
the number of vertices of the graph, and m(θ) the multiplicity of θ as eigenvalue
of S.
We see that there is a 1-1 correspondence between dependent equiangular
systems of n lines and two-graphs on n vertices, and more precisely between
equiangular systems of n lines spanning Rd (with d < n) and two-graphs on n
vertices such that the smallest eigenvalue has multiplicity n − d.
Thus, in order to find large sets of equiangular lines, one has to find large
graphs where the smallest Seidel eigenvalue has large multiplicity (or, rather,
small comultiplicity).
9.6. EQUIANGULAR LINES 143
9.6.2 Bounds on equiangular sets of lines in Rd or Cd

An upper bound for the size of an equiangular system of lines (and hence an
upper bound for the multiplicity of the smallest Seidel eigenvalue of a graph) is
given by the so-called Absolute bound due to M. Gerzon (cf. [221]):
Theorem 9.6.1 (‘Absolute bound’) The cardinality n of a system of equian-

gular lines in Euclidean space Rd is bounded by 12 d(d + 1).
Proof. Let Xi = xi x⊤ i be the rank 1 matrix that is the projection onto the
line ℓi . Then Xi2 = Xi and

1 if i = j
tr Xi Xj = (x⊤
i xj )
2
=
α2 otherwise.
We prove that the matrices Xi are linearly independent. Since they are sym-
1
metric,
P P that there are at most 2 d(d + 1). So,
that will show suppose
P that
2 2
ci Xi = 0. Then i ci Xi Xj = 0 for each j, so that cj (1
P − α ) +Pα ci = 0
for each j. This means that all cj are equal, and since ci = tr ci Xi = 0,
they are all zero.
In Cd one can study lines (1-spaces) in the same way, choosing a spanning unit
vector in each and agreeing that hxi and hyi make angle φ = arccos α where
α = |x∗ y|. (Here x∗ stands for x⊤ .) The same argument now proves
Proposition 9.6.2 The cardinality n of a system of equiangular lines in Cd is

bounded by d2 .
There are very few systems of lines in Rd that meet the absolute bound, but it
is conjectured that systems of d2 equiangular lines in Cd exist for all d. Such
systems are known for d = 1, 2, 3, 4, 5, 6, 7, 8, 19 ([310, 198, 199, 163, 9]). In
quantum information theory they are known as SICPOVMs.
The special bound gives an upper bound for n in terms of the angle φ, or an
upper bound for φ (equivalently, a lower bound for α = cos φ) in terms of n.
Proposition 9.6.3 (‘Special bound’) If there is a system of n > 1 lines in Rd

or Cd such that the cosine of the angle between any two lines is at most α, then
α2 ≥ (n − d)/(n − 1)d, or, equivalently, n ≤ d(1 − α2 )/(1 − α2 d) if 1 − α2 d > 0.
Proof. Let xi (1 ≤ i ≤ n) beP unit vectors in Rd or Cd with |x∗i xj | ≤ α for

i 6= j. Put Xi = xi xi and Y = i Xi − nd I. Then tr Xi Xj = |x∗i xj |2 ≤ α2 for
∗
2
i 6= j, and tr Xi = tr Xi2 = 1. Now 0 ≤ tr Y Y ∗ ≤ n(n − 1)α2 + n − nd .
Complex systems of lines with equality in the special bound are known as equian-
gular tight frames. There is a lot of recent literature.
If equality holds in the absolute bound, then the Xi span the vector space of
all symmetric matrices, and in particular I is a linear combination of the Xi . If
equality holds in the special bound, the same conclusion follows. In both cases
the following proposition shows (in the real case) that the graph Γ belongs to a
regular two-graph.
Proposition 9.6.4 Suppose xi (1 ≤ i ≤ n) are unit vectors in Rd or Cd with

|x∗i xj | = α for i 6= j, where 0 < α <P1. Put Xi = xi x∗i and suppose that
there are constants ci such that I = ci Xi . Then ci = d/n for all i and
n = d(1 − α2 )/(1 − α2 d).
If the xi are vectors in Rd , and G is the Gram matrix of the xi , and G =
I + αS, then S has eigenvalues (n − d)/(αd) and −1/α with multiplicities d and
n − d, respectively. If n > d + 1 and n 6= 2d, then these eigenvalues are odd
integers.
P P
Proof.
P If I = ci Xi then Xj = i ci Xi Xj for each j, so that cj (1 −Pα2 ) +
α2 P ci = 1 for each j. This means that all cj are equal, and since ci =
tr ci Xi = tr I = d, they all equal d/n. Our equation now becomes (d/n)(1 −
α2 ) + α2 d = 1, so that n = d(1 − α2 )/(1 − α2 d).
⊤
If F is the d
⊤
P× n matrix
⊤
Pwhose columns are the vectors xi ,⊤then G = F F ,
while F F = xi xi = Xi = (n/d)I. It follows that F F has eigenvalue
n/d with multiplicity d, and G = F ⊤ F has the same eigenvalues, and in addition
0 with multiplicity n − d. The spectrum of S follows. If the two eigenvalues
of the integral matrix S are not integers, they are conjugate algebraic integers,
and then have the same multiplicity, so that n = 2d. Since S = J − I − 2A, the
eigenvalues of S, when integral, are odd.
Graphs for which the Seidel adjacency matrix S has only two eigenvalues are
strong (cf. §9.1, Proposition 9.1.1) and belong to the switching class of a regular
two-graph (Theorem 9.3.1).
The known lower and upper bounds for the maximum number of equiangular
lines in Rd are given in the table below. For these bounds, see van Lint &
Seidel [229], Lemmens & Seidel [221], Seidel [277] (p. 884).
d 1 2 3 4 5 6 7–14 15 16 17–18
Nmax 1 3 6 6 10 16 28 36 40 48
d 19 20 21 22 23–42 43
Nmax 72–76 90–96 126 176 276 344
Bounds for the size of systems of lines in Rd or Cd with only a few different,
specified, angles, or just with a given total number of different angles, were given
by Delsarte, Goethals & Seidel [129].
9.6.3 Bounds on sets of lines with few angles and sets of

vectors with few distances
In the case of equiangular lines the absolute value of the inner product took
only one value. Generalizing that, one has
Theorem 9.6.5 ([129]) For a set of n unit vectors in Rd such that the abso-
lute value of the inner product between
distinct vectors takes s distinct values
different from 1, one has n ≤ d+2s−1
d−1 . If one of the inner products is 0, then
d+2s−2

n ≤ d−1 .
There are several examples of equality. For example, from the root system
of E8 one gets 120 lines in R8 with |α| ∈ {0, 12 }.
9.6. EQUIANGULAR LINES 145
Theorem 9.6.6 ([129]) For a set of n unit vectors in Cd such that the abso-
lute value of the inner product between
2 distinct vectors takes s distinct values
different from 1, one has n ≤ d+s−1
d−1 . If one of the inner products is 0, then

d+s−1 d+s−2

n ≤ d−1 d−1 .
For example, there are systems of 40 vectors in C4 with |α| ∈ {0, √13 } and
126 vectors in C6 with |α| ∈ {0, 21 }.
For sets of unit vectors instead of sets of lines it may be more natural to look
at the inner product itself, instead of using the absolute value.
Theorem 9.6.7 ([130]) For a set of n unit vectors in Rd such that the inner
product between distinct vectors takes s distinct values, one has n ≤ d+s−1
d−1 +
d+s−2
d+s−2

d−1 . If the set is antipodal, then n ≤ 2 d−1 .
For example, in the antipodal case the upper bound is met with equality for
s = 1 by a pair of vectors ±x (with n = 2), for s = 2 by the vectors ±ei of a
coordinate frame (with n = 2d), and for s = 6 by the set of shortest nonzero
vectors in the Leech lattice in R24 (with inner products −1, 0, ± 14 , ± 12 and size

n = 2 28
5 ).
In the general case the upper bound is met with equality for s = 1 by a
simplex (with n = d + 1). For s = 2 one has
d 2 5 6 22 23 3, 4, 7–21, 24–39
1
Nmax 5 16 27 275 276–277 2 d(d + 1)
with examples of equality in the bound n ≤ 21 d(d + 3) for d = 2, 6, 22. The
upper bounds for d > 6, d 6= 22 are due to Musin [248].
Corollary 9.6.8 ([130]) Let Γ be a regular graph on n vertices, with smallest

eigenvalue θmin < −1 of multiplicity n − d. Then n ≤ 21 d(d + 1) − 1.
(Earlier we saw for strongly regular graphs that n ≤ 12 f (f + 3). Here d =

f + 1, so this gives the same bound, but applies to a larger class of graphs.)
Theorem 9.6.9 ([27]) A set of vectors in Rd such that the distance between
distinct vectors takes s values has size at most d+s
d .
For d ≤ 8, the maximal size of a 2-distance set in Rd was determined by

Lisoněk [230]. The results are
d 1 2 3 4 5 6 7 8
Nmax 3 5 6 10 16 27 29 45

so that equality holds in the Blokhuis bound d+2
2 for d = 1 and d = 8.
The above gave generalizations of the absolute bound. There are also analogs
of the special bound, see [129, 130].
Chapter 10
Association schemes
10.1 Definition
An association scheme with d classes is a finite set X together with d+1 relations
Ri on X such that
(i) {R0 , R1 , . . . , Rd } is a partition of X × X;
(ii) R0 = {(x, x)|x ∈ X};
(iii) if (x, y) ∈ Ri , then also (y, x) ∈ Ri , for all x, y ∈ X and i ∈ {0, ..., d};
(iv) for any (x, y) ∈ Rk the number pkij of z ∈ X with (x, z) ∈ Ri and (z, y) ∈
Rj depends only on i, j and k.
The numbers pkij are called the intersection numbers of the association scheme.
The above definition is the original definition of Bose & Shimamoto [34]; it is
what Delsarte [127] calls a symmetric association scheme. In Delsarte’s more
general definition, (iii) is replaced by:
(iii’) for each i ∈ {0, ..., d} there exists a j ∈ {0, ..., d} such that (x, y) ∈ Ri
implies (y, x) ∈ Rj ,
(iii”) pkij = pkji , for all i, j, k ∈ {0, ..., d}.
It is also very common to require just (i), (ii), (iii’), (iv), and to call the scheme
‘commutative’ when it also satisfies (iii”). Define n = |X|, and ni = p0ii . Clearly,
for each i ∈ {1, ..., d}, (X, Ri ) is a simple graph which is regular of degree ni .
Theorem 10.1.1 The intersection numbers of an association scheme satisfy
(i) pk0j = δjk , p0ij = δij nj , pkij = pkji ,

P P
(ii) i pkij = nj , j nj = n,
(iii) pkij nk = pjik nj ,

P P
(iv) l plij pm
kl = l plkj pm
il .
147
148 CHAPTER 10. ASSOCIATION SCHEMES
Proof. Equations (i), (ii) and (iii) are straightforward. The expressions at
both sides of (iv) count quadruples (w, x, y, z) with (w, x) ∈ Ri , (x, y) ∈ Rj ,
(y, z) ∈ Rk , for a fixed pair (w, z) ∈ Rm .
It is convenient to write the intersection numbers as entries of the so-called
intersection matrices L0 , . . . , Ld :
(Li )kj = pkij .

P k
Note that L0 = I and Li Lj = pij Lk . From the definition it is clear that
an association scheme with two classes is the same as a pair of complemen-
tary strongly regular graphs. If (X, R1 ) is strongly regular with parameters
(v, k, λ, µ), then the intersection matrices of the scheme are
   
0 k 0 0 0 v−k−1
L1 =  1 λ k − λ − 1  , L2 =  0 k − λ − 1 v − 2k + λ  .
0 µ k−µ 1 k−µ v − 2k + µ − 2
10.2 The Bose-Mesner algebra

The relations Ri of an association scheme are described by their adjacency
matrices Ai of order n defined by

1 whenever (x, y) ∈ Ri ,
(Ai )xy =
0 otherwise.
In other words, Ai is the adjacency matrix of the graph (X, Ri ). In terms of

the adjacency matrices, the axioms (i)–(iv) become
Pd
(i) i=0 Ai = J,
(ii) A0 = I,
(iii) Ai = A⊤
i , for all i ∈ {0, . . . , d},
P
(iv) Ai Aj = k pkij Ak , for all i, j, k ∈ {0, . . . , d}.
From (i) we see that the (0, 1) matrices Ai are linearly independent, and by use
of (ii)–(iv) we see that they generate a commutative (d + 1)-dimensional algebra
A of symmetric matrices with constant diagonal. This algebra was first studied
by Bose & Mesner [33] and is called the Bose-Mesner algebra of the association
scheme.
Since the matrices Ai commute, they can be diagonalized simultaneously
(see Marcus & Minc [239]), that is, there exist a matrix S such that for each
A ∈ A, S −1 AS is a diagonal matrix. Therefore A is semisimple and has a
unique basis of minimal idempotents E0 , . . . , Ed (see Burrow [67]). These are
matrices satisfying
d
X
Ei Ej = δij Ei , Ei = I.
i=0
The matrix n1 J is a minimal idempotent (idempotent is clear, and minimal

follows since rk J = 1). We shall take E0 = n1 J. Let P and n1 Q be the matrices
10.2. THE BOSE-MESNER ALGEBRA 149
relating our two bases for A:

d
X d
1X
Aj = Pij Ei , Ej = Qij Ai .
i=0
n i=0
Then clearly
P Q = QP = nI.
It also follows that
Aj Ei = Pij Ei ,
which shows that the Pij are the eigenvalues of Aj and that the columns of Ei
are the corresponding eigenvectors. Thus mi = rk Ei is the multiplicity of the
eigenvalue Pij of Aj (provided that Pij 6= Pkj for k 6= i). We see that m0 = 1,
P
i mi = n, and mi = trace Ei = n(Ei )jj (indeed, Ei has only eigenvalues 0
and 1, so rk Ek equals the sum of the eigenvalues).
Theorem 10.2.1 The numbers Pij and Qij satisfy
(i) Pi0 = Qi0 = 1, P0i = ni , Q0i = mi ,

Pd
(ii) Pij Pik = l=0 pljk Pil ,
P P
(iii) mi Pij = nj Qji , i mi Pij Pik = nnj δjk , i ni Qij Qik = nmj δjk ,
(iv) |Pij | ≤ nj , |Qij | ≤ mj .

P P
Proof. Part (i) follows easily from i Ei = I = A0 , i Ai = JP= nE0 ,
l
Ai J = ni J, and trace Ei = mi . Part (ii) follows from Aj Ak = l pjk Al .
The first equality in (iii) follows from mi Pij = trAj Ei = nj Qji , and the other
two follow since P Q = nI. The first inequality of (iv) holds because the Pij
are eigenvalues of the nj -regular graphs (X, Rj ). The second inequality then
follows by use of (iii).
Relations (iii) are often referred to as the orthogonality relations, since they
state that the rows (and columns) of P (and Q) are orthogonal with respect to
a suitable weight function.
If d = 2, and (X, R1 ) is strongly regular with parameters (v, k, λ, µ) and
spectrum k 1 rf sg , the matrices P and Q are
   
1 k v−k−1 1 f g
P = 1 r −r − 1  , Q =  1 f r/k gs/k .
r+1 s+1
1 s −s − 1 1 −f v−k−1 −g v−k−1
In general the matrices P and Q can be computed from the intersection

numbers of the scheme, as follows from the following
Theorem 10.2.2 For i = 0, . . . , d, the intersection matrix Lj has eigenvalues

Pij (0 ≤ i ≤ d).
Proof. Theorem 10.2.1(ii) yields

X X
Pil (Lj )lk (P −1 )km = Pij Pik (P −1 )km = δim Pij ,
k,l k
hence P Lj P −1 = diag (P0j , . . . , Pdj ).

1 −1
Thanks to this theorem, it is relatively easy to compute P , Q (= nP ) and
mi (= Q0i ). It is also possible to express P and Q in terms of the (common)
eigenvectors of the Lj . Indeed, P Lj P −1 = diag (P0j , . . . , Pdj ) implies that the
rows of P are left eigenvectors and the columns of Q are right eigenvectors.
In particular, mi can be computed from the right eigenvector ui and the left
eigenvector vi⊤ , normalized such that (ui )0 = (vi )0 = 1, by use of mi u⊤ i vi = n.
Clearly, each mi must be an integer. These are the rationality conditions for
an association scheme. As we saw in the case of a strongly regular graph, these
conditions can be very powerful.
10.3 The Linear Programming Bound

One of the main reasons association schemes have been studied is that they
yield upper bounds for the size of substructures.
Let Y be a nonempty subset of X, and let its inner distribution be the
vector a defined by ai = |(Y × Y ) ∩ Ri |/|Y |, the average number of elements of
Y in relation Ri to a given one. Let χ be the characteristic vector of Y . Then
ai = |Y1 | χ⊤ Ai χ.
Theorem 10.3.1 (Delsarte) aQ ≥ 0.

P P
Proof. We have |Y |(aQ)j = |Y | ai Qij = χ⊤ Qij Ai χ = nχ⊤ Ej χ ≥ 0
since Ej is positive semidefinie.
Example Consider the schemes of the triples from a 7-set, where two triples
are in relation Ri when they have 3 − i elements in common (i = 0, 1, 2, 3). We
find
   
1 12 18 4 1 6 14 14
 1 5 −3 −3   0 −7/2 
P =  and Q =  1 5/2 .
 1 0 −3 2   1 −1 −7/3 7/3 
1 −3 3 −1 1 −9/2 7 −7/2
How many triples can we find such that any two meet in at most one point? For
the inner distribution a of such a collection Y we have a1 = 0, so a = (1, 0, s, t),
and aQ ≥ 0 gives the three inequalities
6 − s − 92 t ≥ 0, 14 − 73 s + 7t ≥ 0, 14 + 73 s − 72 t ≥ 0.
The linear programming problem is to maximize |Y | = 1 + s + t given these

inequalities, and the unique solution is s = 6, t = 0. This shows that one can
have at most 7 triples that pairwise meet in at most one point in a 7-set, and if
one has 7, then no two are disjoint. Of course an example is given by the Fano
plane.
How many triples can we find such that any two meet in at least one point?
Now a = (1, r, s, 0) and the optimal solution of aQ ≥ 0 is (1, 8, 6, 0). An example
of such a collection is given by the set of 15 triples containing a fixed point.
How many triples can we find such that no two meet in precisely one point?
Now a = (1, r, 0, t) and the maximum value of 1 + r + t is 5. An example is
given by the set of 5 triples containing two fixed points.
10.4. THE KREIN PARAMETERS 151
10.4 The Krein parameters

The Bose-Mesner algebra A is not only closed under ordinary matrix multiplica-
tion, but also under componentwise (Hadamard, Schur) multiplication (denoted
◦). Clearly {A0 , . . . , Ad } is the basis of minimal idempotents with respect to
this multiplication. Write
d
1X k
Ei ◦ Ej = qij Ek .
n
k=0
k k
The numbers qij thus defined are called the Krein parameters. (Our qij are
those of Delsarte, but differ from Seidel’s [276] by a factor n.) As expected, we
now have the analogue of Theorems 10.1.1 and 10.2.1.
Theorem 10.4.1 The Krein parameters of an association scheme satisfy
k 0 k k
(i) q0j = δjk , qij = δij mj , qij = qji ,
P k P
(ii) i qij = mj , j mj = n,
k j
(iii) qij mk = qik mj ,
P l m P l m
(iv) l qij qkl = l qkj qil ,
Pd l
(v) Qij Qik = l=0 qjk Qil ,
k
P
(vi) nmk qij = l nl Qli Qlj Qlk .
P
Proof. P
P Let (A) denote the sum of all Pentries of the matrix A. Then JAJ =
(A)J, (A ◦ B) = trace AB ⊤ and (Ei ) = 0 if i 6= 0, since P then Ei J =
nEi E0 = 0. Now (i) follows by use of Ei ◦ E0 = n1 Ei , qij 0
= (Ei ◦ Ej ) =
trace Ei Ej = δij mj , and Ei ◦ Ej = Ej ◦ Ei , respectively. Equation (iv) follows
by evaluating Ei ◦ Ej ◦ Ek in two ways, and (iii) follows from (iv) by taking
m = 0. Equation (v) follows from evaluating Ai ◦ EjP◦ Ek in two ways, and
(vi) follows from (v), using the orthogonality relation l nl Qlj Qlk = δmk mk n.
Finally, by use of (iii) we have
X X j X
k
mk qij = qik mj = n · trace (Ei ◦ Ek ) = n (Ei )ll (Ek )ll = mi mk ,
j j l
proving (ii).
The above results illustrate a dual behavior between ordinary multiplication,
the numbers pkij and the matrices Ai and P on the one hand, and Schur mul-
k
tiplication, the numbers qij and the matrices Ei and Q on the other hand. If
two association schemes have the property that the intersection numbers of one
are the Krein parameters of the other, then the converse is also true. Two such
schemes are said to be (formally) dual to each other. One scheme may have
several (formal) duals, or none at all (but when the scheme is invariant under a
regular abelian group, there is a natural way to define a dual scheme, cf. Del-
sarte [127]). In fact usually the Krein parameters are not even integers. But
they cannot be negative. These important restrictions, due to Scott [270] are
the so-called Krein conditions.
k
Theorem 10.4.2 The Krein parameters of an association scheme satisfy qij ≥
0 for all i, j, k ∈ {0, . . . , d}.
Proof. The numbers n1 qij k

(0 ≤ k ≤ d) are the eigenvalues of Ei ◦ Ej (since
1 k
(Ei ◦ Ej )Ek = n qij Ek ). On the other hand, the Kronecker product Ei ⊗ Ej
is positive semidefinite, since each Ei is. But Ei ◦ Ej is a principal submatrix
of Ei ⊗ Ej , and therefore is positive semidefinite as well, i.e., has no negative
eigenvalue.
The Krein parameters can be computed by use of equation 10.4.1(vi). This
equation also shows that the Krein condition is equivalent to
X
nl Qli Qlj Qlk ≥ 0 for all i, j, k ∈ {0, . . . , d}.
l
In case of a strongly regular graph we obtain

1 f2 r3 (r + 1)3
q11 = 1+ 2 − ≥ 0,
v k (v − k − 1)2

2 g2 s3 (s + 1)3
q22 = 1+ 2
− ≥0
v k (v − k − 1)2
(the other Krein conditions are trivially satisfied in this case), which is equivalent
to the result mentioned in section §8.1.5.
Neumaier [250] generalized Seidel’s absolute bound to association schemes,
and obtained the following.
Theorem 10.4.3 The multiplicities mi (0 ≤ i ≤ d) of an association scheme

with d classes satisfy
X
mi m j if i 6= j,
mk ≤ 1
k
m
2 i (m i + 1) if i = j.
qij 6=0
Proof. The left hand side equals rk (Ei ◦Ej ). But rk (Ei ◦Ej ) ≤ rk (Ei ⊗Ej ) =
rk Ei · rk Ej = mi mj . And if i = j, then rk (Ei ◦ Ei ) ≤ 12 mi (mi + 1). Indeed,
if the rows of Ei are linear combinations of mi rows, then the rows of Ei ◦ Ei
are linear combinations of the mi + 21 mi (mi − 1) rows that are the elementwise
products of any two of these mi rows.
1
For strongly regular graphs with q11 = 0 we obtain Seidel’s bound: v ≤ 12 f (f +
1
3). But in case q11 > 0, Neumaier’s result states that the bound can be improved
to v ≤ 12 f (f + 1).
10.5 Automorphisms
Let π be an automorphism of an association scheme, and suppose there are Ni
points x such that x and π(x) are in relation Ri .
1
Pd
Theorem 10.5.1 (G. Higman) For each j the number n i=0 Ni Qij is an al-
gebraic integer.
10.6. P - AND Q-POLYNOMIAL ASSOCIATION SCHEMES 153
Proof. The automorphism is represented by a permutation matrix S where

SM = M S for each M in the Bose-Mesner algebra. Let E = Ej be one of the
idempotents. Then E has eigenvalues 0 and 1, and S has eigenvalues that are
roots of unity, so ES has eigenvalues that
P are zero or a root of unity, and trES
is an algebraic integer. But trES = n1 i Ni Qij .
P P
If one puts aj = n1 i Ni Qij , then Nh = j aj Pjh for all h.
10.5.1 The Moore graph on 3250 vertices

Let Γ be a strongly regular graph with parameters (v, k, λ, µ) = (3250, 57, 0, 1)
(an unknown Moore graph of diameter two, cf. Theorem 8.1.5).
 
1 1729 1520
637
For such a graph Q =  1 3 − 640
3
.
13 10
1 −3 3
Aschbacher [10] proved that there is no such graph with a rank three group.
G. Higman (unpublished, cf. Cameron [75]) proved that there is no such graph
with a vertex transitive group.
Proposition 10.5.2 (G. Higman) Γ is not vertex-transitive.
Proof. Consider any nontrivial group of automorphisms G of such a graph.

The collection of points fixed by G has the properties λ = 0 and µ = 1. Also,
two nonadjacent fixed vertices are adjacent to the same number of fixed vertices,
so the fixed subgraph is either a strongly regular Moore graph (and then has
5, 10 or 50 vertices), or all fixed vertices have distance at most 1 to some fixed
vertex (so that there are at most k + 1 = 58 of them).
Consider an involution π. If π does not interchange the endpoints of some
edge, then N1 = 0, and N0 + N2 = 3250. But if {x, y} is an orbit of π, then
the unique common neighbor z of x and y is fixed, and z occurs for at most
28 pairs {x, y}, so N2 ≤ 56N0 , so that N0 = 58, N1 = 0, N2 = 3192 and
1 13 133
3250 (58 × 1729 − 3192 × 3 ) = 5 is not an integer, contradiction.
So, π must interchange two adjacent points x and y, and hence interchanges
the remaining 56 neighbors u of x with the remaining 56 neighbors v of y. If
{u, v} is such an orbit, then the unique common neighbor of u and v is fixed,
and these are all the fixed points. So N0 = 56, that is, π is an odd permutation,
since it is the product of 1597 transpositions. Let N be the subgroup of G
consisting of the even permutations. Then N does not have any involutions,
so is not transitive, and if G was transitive N has two orbits interchanged by
any element outside N . But α has fixed points and cannot interchange the two
orbits of N . A contradiction, so G was not transitive.
10.6 P - and Q-polynomial association schemes

In many cases, the association scheme carries a distance function such that rela-
tion Ri is the relation of having distance i. Such schemes are called metric. They
are characterized by the fact that pijk is zero whenever one of i, j, k is larger than
the sum of the other two, while pijk is nonzero for i = j + k. Note that whether a
scheme is metric depends on the ordering of the relations Ri . A scheme may be
metric for more than one ordering. Metric association schemes are essentially
the same objects as distance-regular graphs (see Chapter 11 below).
i i
Dually, a cometric scheme is defined by qjk = 0 for i > j + k and qjk > 0 for
i = j + k.
There are several equivalent formulations of the metric (cometric) property.
An association scheme is called P -polynomial if there exist polynomials fk
of degree k with real coefficients, and real numbers zi such that Pik = fk (zi ).
Clearly we may always take zi = Pi1 . By the orthogonality relation 10.2.1(iii)
we have X X
mi fj (zi )fk (zi ) = mi Pij Pik = nnj δjk ,
i i
which shows that the fk are orthogonal polynomials.

Dually, a scheme is called Q-polynomial when the same holds with Q instead
of P . The following result is due to Delsarte [127] (Theorem 5.6, p. 61).
Theorem 10.6.1 An association scheme is metric (resp. cometric) if and only

if it is P -polynomial (resp. Q-polynomial).
Proof. Let the scheme be metric. Then
A1 Ai = pi−1 i i+1
1i Ai−1 + p1i Ai + p1i Ai+1 .
Since pi+1
1i 6= 0, Ai+1 can be expressed in terms of A1 , Ai−1 and Ai . Hence for
each j there exists a polynomial fj of degree j such that Aj = fj (A1 ), and it
follows that Pij Ei = Aj Ei = fj (A1 )Ei = fj (Pi1 )Ei , hence Pij = fj (Pi1 ).
Now suppose that the scheme is P -polynomial. Then the fj are orthog-
onal polynomials, and therefore they satisfy a 3-term recurrence relation (see
Szegö [291] p. 42)
αj+1 fj+1 (z) = (βj − z)fj (z) + γj−1 fj−1 (z).
Hence
Pi1 Pij = −αj+1 Pi,j+1 + βj Pij + γj−1 Pi,j−1 for i = 0, . . . , d.

P l l
Since Pi1 Pij = l p1j Pil and P is nonsingular, it follows that p1j = 0 for
|l − j| > 1. Now the full metric property easily follows by induction. The proof
for the cometric case is similar.
Given a sequence of nonzero real numbers, let its number of sign changes be
obtained by first removing all zeros from the sequence, and then counting the
number of consecutive pairs of different sign. (Thus, the number of sign changes
in 1, −1, 0, 1 is 2.)
Proposition 10.6.2 (i) Let (X, R) be a P -polynomial association scheme, with

relations ordered according to the P -polynomial ordering and eigenspaces ordered
according to descending real order on the θi := Pi1 . Then both row i and column
i of both matrices P and Q have precisely i sign changes (0 ≤ i ≤ d).
(ii) Dually, if (X, R) is a Q-polynomial association scheme, and the eigenspaces
are ordered according to the Q-polynomial ordering and the relations are ordered
according to descending real order on the σi := Qi1 , then row i and column i of
the matrices P and Q have precisely i sign changes (0 ≤ i ≤ d).
10.7. EXERCISES 155
Proof. Since mi Pij = nj Qji the statements about P and Q are equivalent.
Define polynomials pj of degree j for 0 ≤ j ≤ d + 1 by p−1 (x) = 0, p0 (x) = 1,
(x − aj )pj (x) = bj−1 pj−1 + cj+1 pj+1 (x), taking cd+1 = 1. Then Aj = pj (A),
and pd+1 (x) = 0 has as roots the eigenvalues of A. The numbers in row j of P
are pi (θj ) (0 ≤ i ≤ d), and by the theory of Sturm sequences the number of sign
changes is the number of roots of pd+1 larger than θj , which is j. The numbers
in column i of P are the values of pi evaluated at the roots of pd+1 . Since pi
has degree i, and there is at least one root of pd+1 between any two roots of pi
there are i sign changes. The proof in the Q-polynomial case is similar.
Example Consider the Hamming scheme H(4, 2), the association scheme on
the binary vectors of length 4, where the relation is their Hamming distance.
Now
 
1 4 6 4 1
 1 2 0 −2 −1 
 
P =Q= 1  0 −2 0 1 
.
 1 −2 0 2 −1 
1 −4 6 −4 1
10.7 Exercises
Exercise 1 Show that the number of relations of valency 1 in an association
scheme is 2m for some m ≥ 0, and 2m |n. Hint: the relations of valency 1 form
an elementary abelian 2-group with operation i ⊕ j = k when Ai Aj = Ak .
Exercise 2 Show that for the special case where Y is a coclique in a strongly
regular graph, the linear programming bound is the Hoffman bound (Theo-
rem 4.1.2).
Exercise 3 Show that if Γ is a relation of valency k in an association scheme,
and θ is a negative eigenvalue of Γ, then |S| ≤ 1 − k/θ for each clique S in Γ.
Exercise 4 Consider a primitive strongly regular graph Γ on v vertices with
eigenvalues k 1 , rf , sg (k > r > s) with a Hoffman coloring (that is a coloring
with 1 − k/s colors). Consider the following relations on the vertex set of Γ:
R0 : identity,
R1 : adjacent in Γ,
R2 : nonadjacent in Γ with different colors,
R3 : nonadjacent in Γ with the same color.
Prove that these relations define an association scheme on the vertex set of Γ,
and determine the matrices P and Q.
Chapter 11
Distance regular graphs
Consider a connected simple graph with vertex set X of diameter d. Define

Ri ⊂ X 2 by (x, y) ∈ Ri whenever x and y have graph distance i. If this defines
an association scheme, then the graph (X, R1 ) is called distance-regular. By the
i+j
triangle inequality, pkij = 0 if i + j < k or |i − j| > k. Moreover, pij > 0.
Conversely, if the intersection numbers of an association scheme satisfy these
conditions, then (X, R1 ) is easily seen to be distance-regular.
Many of the association schemes that play a rôle in combinatorics are met-
ric. Families of distance-regular graphs with unbounded diameter include the
Hamming graphs, the Johnson graphs, the Grassmann graphs and graphs asso-
ciated to dual polar spaces. Recently Van Dam & Koolen [121] constructed a
new such family, the 15th, and the first without transitive group.
Many constructions and results for strongly regular graphs are the d = 2
special case of corresponding results for distance-regular graphs.
The monograph [48] is devoted to the theory of distance-regular graphs, and
gives the state of the theory in 1989.
11.1 Parameters
Conventionally, the parameters are bi = pii+1,1 and ci = pii−1,1 (and ai = pii,1 ).
The intersection array of a distance-regular graph of diameter d is {b0 , . . . , bd−1 ;
c1 , . . . , cd }. The valencies p0i,i , that were called ni above, are usually called ki
here. We have ci ki = bi−1 ki−1 . The total number of vertices is usually called v.
It is easy to see that one has b0 ≥ b1 ≥ . . . ≥ bd−1 and c1 ≤ c2 ≤ . . . ≤ cd
and cj ≤ bd−j (1 ≤ j ≤ d).
11.2 Spectrum
A distance-regular graph Γ of diameter d has d + 1 distinct eigenvalues, and
the spectrum is determined by the parameters. (Indeed, the matrices P and
Q of any association scheme are determined by the parameters pijk , and for a
distance-regular graph the pijk are determined again in terms of the bi and ci .)
The eigenvalues of Γ are the eigenvalues of the tridiagonal matrix L1 = (pj1k )
157
158 CHAPTER 11. DISTANCE REGULAR GRAPHS
of order d + 1 that here gets the form

 
0 b0 0
 c 1 a1 b1 
 
L1 =  c2 a2 b2 .

 ... ... 
0 cd ad
If L1 u = θu
Pand u0 = 1, then the multiplicity of θ as eigenvalue of Γ equals
m(θ) = v/( ki u2i ).
11.3 Examples
11.3.1 Hamming graphs
Let Q be a set of size q. The Hamming graph H(d, q) is the graph with vertex
set Qd , where two vertices are adjacent when they agree in d − 1 coordinates.
This graph is distance-regular, with parameters ci = i,bi = (q − 1)(d − i),
diameter d and eigenvalues (q − 1)d− qi with multiplicity di (q − 1)i (0 ≤ i ≤ d).
(Indeed, H(d, q) is the Cartesian product of d copies of Kq , see §1.4.6.)
For q = 2 this graph is also known as the hypercube 2d , often denoted Qd .
For d = 2 the graph H(2, q) is also called L2 (q).
Cospectral graphs
In §1.8.1 we saw that there are precisely two graphs with the spectrum of H(4, 2).
In §8.2 we saw that there are precisely two graphs with the spectrum of H(2, 4).
Here we give a graph cospectral with H(3, 3) (cf. [183]).
The graphs H(d, q) have q d vertices, and dq d−1 maximal cliques (‘lines’) of size
u u u
u

u u

u
u
u

u u u
u
u
u

u
u
u

u u u
u
u
u

u
u
u
Figure 11.1: The geometry of the Hamming graph H(3, 3)

q. Let N be the point-line incidence matrix. Then N N ⊤ − dI is the adjacency
matrix of Γ = H(d, q), and N ⊤ N −qI is the adjacency matrix of the graph ∆ on
the lines, where two lines are adjacent when they have a vertex in common. It
follows that for d = q the graphs Γ and ∆ are cospectral. In Γ any two vertices
at distance two have c2 = 2 common neighbors. If q ≥ 3, then two vertices at
distance two in ∆ have 1 or q common neighbors (and both occur), so that ∆
is not distance-regular, and in particular not isomorphic to Γ. For q = 3 the
geometry is displayed in Figure 11.1. See also §13.2.2.
11.3. EXAMPLES 159
11.3.2 Johnson graphs
Let X be a set of size n. The Johnson graph J(n, m) is the graph with vertex
set Xm , the set of all m-subsets of X, where two m-subsets are adjacent when
they have m − 1 elements in common. For example, J(n, 0) has a single vertex;
J(n, 1) is the complete graph Kn ; J(n, 2) is the triangular graph T (n).
This graph is distance-regular, with parameters ci = i2 , bi = (m − i)(n −
m − i), diameter
d =nmin(m,
n − m) and eigenvalues (m − i)(n − m − i) − i with
multiplicity ni − i−1 .

The Kneser graph K(n, m) is the graph with vertex set X m , where two
m-subsets are adjacent when they have maximal distance in J(n, m) (i.e., are
disjoint when n ≥ 2m, and have 2m − n elements in common otherwise). These
graphs are not distance-regular in general, but the Odd graph Om+1 , which
equals K(2m + 1, m), is.
Sending a vertex (m-set) to its complement in X is an isomorphism from
J(n, m) onto J(n, n − m) and from K(n, m) onto K(n, n − m). Thus, we may
always assume that n ≥ 2m.
11.3.3 Grassmann graphs
Let V be a vector space of dimension n over

the field Fq . The Grassmann graph
V
Gr(n, m) is the graph with vertex set m , the set of all m-subspaces of V , where
two m-subspaces are adjacent when they intersect in an (m − 1)-space. This
2 n−m−i
graph is distance-regular, with parameters ci = 1i , bi = q 2i+1 m−i ,

i+1 m−i n−m−i
1 i 1
diameter d = min(m,
n n − m), and
eigenvalues q 1 1 − 1 with
multiplicity ni − i−1 . (Here ni = (q n − 1) · · · (q n−i+1 − 1)/(q i − 1) · · · (q − 1)
is the q-binomial coefficient, the number of m-subspaces of an n-space.)
11.3.4 Van Dam-Koolen graphs
Van Dam & Koolen [121] construct distance-regular graphs vDK(m) with the
same parameters as Gr(2m + 1, m). (They call them the twisted Grassmann
graphs.) These graphs are ugly, the group of automorphisms is not transi-
tive. The existence of such examples reinforces the idea that the parameters of
distance-regular graphs of large diameter are strongly restricted, while there is
some freedom for the actual structure. The construction is as follows. Let V
be a vector space of dimension 2m + 1 over Fq , and let H be a hyperplane of
V . Take as vertices the (m + 1)-subspaces of V not contained in H, and the
(m − 1)-subspaces contained in H, where two subspaces of the same dimension
are adjacent when their intersection has codimension 1 in both, and two sub-
spaces of different dimension are adjacent when one contains the other. This
graph is the line graph (concurrency graph on the set of lines) of the partial lin-
ear space of which the points are the m-subspaces of V , with natural incidence,
while the point graph (collinearity graph on the set of points) is Gr(2m + 1, m).
It follows that vDK(m) and Gr(2m + 1, m) are cospectral.
11.4 Bannai-Ito conjecture

The most famous problem about distance-regular graphs was the Bannai-Ito
conjecture ([17], p. 237): show that there are only finitely many distance-
regular graphs with fixed valency k larger than 2. After initial work by Bannai
& Ito, the conjecture was attacked by Jack Koolen and coauthors in a long series
of papers. After 25 years a complete proof was given by Sejeong Bang, Arturas
Dubickas, Jack Koolen, and Vincent Moulton [15].
11.5 Connectedness
For strongly regular graphs we had Theorem 8.3.2 stating that the vertex con-
nectivity κ(Γ) equals the valency k. In [55] it was shown that the same holds
for distance-regular graphs.
For strongly regular graphs we also had Proposition 8.3.1 that says that
the induced subgraph on the vertices at maximal distance from a given ver-
tex is connected. This is a very important property, but for distance-regular
graphs additional hypotheses are needed. For example, there are two gener-
alized hexagons with parameters GH(2, 2) (duals of each other) and in one of
them the subgraphs Γ3 (x) are disconnected.
11.6 Degree of eigenvalues

For strongly regular graphs we saw that eigenvalues are integral, except in the
‘half case’ where they are quadratic. Something similar happens for distance-
regular graphs.
Polygons have eigenvalues of high degree: for an n-gon the degree of the i-th
eigenvalue is φ(m) where m = gcd(i, n), where φ is the Euler totient function.
But elsewhere only integral and quadratic eigenvalues seem to occur.
For the case of a P - and Q-polynomial scheme of diameter at least 34, Bannai
& Ito [17] Theorem 7.11 show that the eigenvalues are integers.
There is precisely one known distance-regular graph of valency larger than
2 with a cubic eigenvalue, namely the Biggs-Smith graph, the unique graph
with intersection array {3, 2, 2,√2, 1, 1, 1; 1, 1, 1, 1, 1, 1, 3}. It has 102 vertices,
and spectrum 31 218 017 ((1 ± 17)/2)9 θj16 where the θj are the three roots of
θ3 + 3θ2 − 3 = 0.
A result in this direction is
Proposition 11.6.1 The only distance-regular graph of diameter 3 with a cubic

eigenvalue is the heptagon.
Proof. Let Γ be a distance-regular graph of diameter 3 on n vertices with

a cubic eigenvalue. Since algebraically conjugate eigenvalues have the same
multiplicity we have three eigenvalues θi with multiplicity f = (n − 1)/3. Since
trA = 0 we find that θ1 + θ2 + θ3 = −k/f . Now k/f is rational and an algebraic
integer, hence an integer, and k ≥ (n − 1)/3. The same reasoning applies to
Ai for i = 2, 3 and hence ki ≥ (n − 1)/3, and we must have equality. Since
k = k2 = k3 we see that b1 = c2 = b2 = c3 .
11.7. MOORE GRAPHS AND GENERALIZED POLYGONS 161
Write µ := c2 . The distinct eigenvalues k, θ1 , θ2 , θ3 of A are the eigenvalues

of the matrix L1 (Theorem 10.2.2) and hence k − 1 = k + θ1 + θ2 + θ3 = trL1 =
a1 + a2 + a3 = (k − µ − 1) + (k − 2µ) + (k − µ), so that k = 2µ and a2 = 0.
Let d(x, y) = 3 and put A = Γ(x) ∩ Γ2 (y), B = Γ2 (x) ∩ Γ(y), so that
|A| = |B| = c3 = µ. Every vertex in B is adjacent to every vertex in A, and
hence two vertices in B have at least µ + 1 common neighbors, so must be
adjacent. Thus B is a clique, and µ = |B| ≤ a2 + 1, that is, µ = 1, k = 2.
11.7 Moore graphs and generalized polygons

Any k-regular graph of diameter d has at most
1 + k + k(k − 1) + . . . + k(k − 1)d−1
vertices, as is easily seen. A graph for which equality holds is called a Moore
graph. Moore graphs are distance-regular, and those of diameter 2 were dealt
with in Theorem 8.1.5. Using the rationality conditions Damerell [123] and
Bannai & Ito [16] showed:
Theorem 11.7.1 A Moore graph with diameter d ≥ 3 is a (2d + 1)-gon.
A strong non-existence result of the same nature is the theorem of Feit & G. Hig-
man [139] about finite generalized polygons. We recall that a generalized m-gon
is a point-line incidence geometry such that the incidence graph is a connected,
bipartite graph of diameter m and girth 2m. It is called regular of order (s, t)
for certain (finite or infinite) cardinal numbers s, t if each line is incident with
s + 1 points and each point is incident with t + 1 lines. From such a regular
generalized m-gon of order (s, t), where s and t are finite and m ≥ 3, we can
construct a distance-regular graph with valency s(t + 1) and diameter d = ⌊ m 2⌋
by taking the collinearity graph on the points.
Theorem 11.7.2 A finite generalized m-gon of order (s, t) with s > 1 and
t > 1 satisfies m ∈ {2, 3, 4, 6, 8}.
Proofs of this theorem can be found in Feit & Higman [139], Brouwer, Cohen
& Neumaier [48] and Van Maldeghem [298]; again the rationality conditions do
the job. The Krein conditions yield some additional information:
Theorem 11.7.3 A finite regular generalized m-gon with s > 1 and t > 1
satisfies s ≤ t2 and t ≤ s2 if m = 4 or 8; it satisfies s ≤ t3 and t ≤ s3 if m = 6.
This result is due to Higman [191] and Haemers & Roos [181].
11.8 Primitivity
A distance-regular graph Γ of diameter d is called imprimitive when one of the
relations (X, Ri ) with i 6= 0 is disconnected. This can happen in three cases:
either Γ is an n-gon, and i|n, or i = 2, and Γ is bipartite, or i = d, and Γ is
antipodal, that is, having distance d is an equivalence relation. Graphs can be
both bipartite and antipodal. The 2n-gons fall in all three cases.
11.9 Euclidean representations

Let Γ be distance regular, and let θ be a fixed eigenvalue. Let E = Ej be the
idempotent in the association
P scheme belonging to θ, so that AE = θE. Let
ui = Qij /n, so that E = ui Ai . Let f = rkE.
The map sending vertex x of Γ to the vector x̄ = Eex , column x of E,
provides a representation of Γ by vectors in an f -dimensional Euclidean space,
namely the column span of E, where graph distances are translated into inner
products: if d(x, y) = i then (x̄, ȳ) = Exy = ui .
If this map is not injective, and x̄ = ȳ for two vertices x, y at distance i 6= 0,
then ui = u0 and any two vertices at distance i have the same image. For i = 1
this happens when θ = k. Otherwise, Γ is imprimitive, and either i = 2 and Γ
is bipartite and θ = −k, or i = d and Γ is antipodal, or 2 < i < d and Γ is a
polygon.
This construction allows one to translate problems about graphs into prob-
lems in Euclidean geometry. Especially when f is small, this is a very useful
tool.
As an example of the use of this representation, let us prove Terwilliger’s Tree
Bound. Call an induced subgraph T of Γ geodetic when distances measured in
T equal distances measured in Γ.
Proposition 11.9.1 Let Γ be distance regular, and let θ be an eigenvalue dif-

ferent from ±k. Let T be a geodetic tree in Γ. Then the multiplicity f of the
eigenvalue θ is at least the number of endpoints of T .
Proof. We show that the span of the vectors x̄ for x ∈ T has a dimension
not less than the number e of endpoints of T . Induction on the size of T . If
T = {x, y} then x̄ 6= ȳ since k 6= θ. Assume |T | > 2. If x ∈ T , and S is
the set of endpoints of T adjacent to x, then for y, z ∈ S and w ∈ T \ S we
have (w̄, ȳ − z̄) = 0. Pick x such that S is nonempty, and x is an endpoint of
T ′ = T \ S. By induction dimhw̄ | w ∈ T ′ i ≥ e − |S| + 1. Since θ 6= ±k we have
dimhȳ − z̄ | x, y ∈ Si = |S| − 1.
Example For a distance-regular graph without triangles, f ≥ k. Equality can
hold. For example, the Higman-Sims graph is strongly regular with parameters
(v, k, λ, µ) = (100, 22, 0, 6) and spectrum 221 277 (−8)22 .
11.10 Extremality
This section gives a simplified account of the theory developed by Fiol and
Garriga and coauthors. The gist is that among the graphs with a given spectrum
with d + 1 distinct eigenvalues the distance-regular graphs are extremal in the
sense that they have a maximal number of pairs of vertices at mutual distance d.
Let Γ be a connected k-regular graph with adjacency matrix A with eigen-
values k = θ1 ≥ · · · ≥ θn . Suppose that A has precisely d+1 distinct eigenvalues
(so that the diameter of Γ is at most d). Define an inner product on the (d + 1)-
dimensional vector space of real polynomials modulo the minimum polynomial
of A by
n
1 1X
hp, qi = tr p(A)q(A) = p(θi )q(θi ).
n n i=1
11.11. EXERCISES 163
Note that hp, pi ≥ 0 for all p, and hp, pi = 0 if and only if p(A) = 0. By applying
Gram-Schmidt to the sequence of polynomials xi (0 ≤ i ≤ d) we find a sequence
of orthogonal polynomials pi of degree i (0 ≤ i ≤ d) satisfying hpi , pj i = 0 for
i 6= j and hpi , pi i = pi (k). This latter normalization is possible since pi (k) 6= 0.
(Indeed, suppose that pi changes sign at values αj (0 ≤ j ≤ h) inside the
Qh
interval (θn , k). Put q(x) = j=1 (x − αj ). Then all terms in hpi , qi have the
same sign, and not all are zero, so hpi , qi 6= 0, hence h = i, so that all zeros of
pi are in the interval (θn , k), and pi (k) 6= 0.)
The Hoffman polynomial (the polynomial p such that p(A) = J) equals
p0 + . . . + pd . Indeed, hpi , pi = n1 tr pi (A)J = pi (k) = hpi , pi i for all i.
If Γ is distance-regular, then the pi are the polynomials for which Ai = pi (A).
Theorem 11.10.1 Let Γ be connected and regular of degree k, with P d+1 distinct
eigenvalues. Define the polynomials pi as above. Let kd := n1 x kd (x) be
the average number of vertices at distance d from a given vertex in Γ. Then
kd ≤ pd (k), and equality holds if and only if Γ is distance-regular.
Proof. We follow Fiol, Gago & Garriga [144]. Use the inner product hM, N i =
1 ⊤
n tr M N on the space MP n (R) of real matrices of order n. If M, N are sym-
metric, then hM, N i = n1 x,y (M ◦ N )xy . If M = p(A) and N = q(A) are
polynomials in A, then hM, N i = hp, qi.
Since hAd , pd (A)i = hAd , Ji = kd , the orthogonal projection A′d of Ad on the
space hI, A, ..., Ad i = hp0 (A), ..., pd (A)i of polynomials in A equals
X hAd , pj (A)i hAd , pd (A)i kd
A′d = pj (A) = pd (A) = pd (A).
j
hpj , pj i pd (k) pd (k)
2
Now ||A′d ||2 ≤ ||Ad ||2 gives kd /pd (k) ≤ kd , and the inequality follows since
pd (k) > 0. When equality holds, Ad = pd (A).
Now it follows
P by downward P induction on h that Ah = ph (A) (0 ≤ h ≤ d).
Indeed, from j pj (A) = J = j Aj it follows that p0 (A) + · · · + ph (A) =
A0 + · · · + Ah . Hence ph (A)xy = 0 if d(x, y) > h, and ph (A)xy = 1 if d(x, y) = h.
Since hxph+1 , pj i = hph+1 , xpj i = 0 for j 6= h, h + 1, h + 2, we have xph+1 =
aph + bph+1 + cph+2 and hence AAh+1 = aph (A) + bAh+1 + cAh+2 for certain
a, b, c with a 6= 0. But then ph (A)xy = 0 if d(x, y) < h, so that ph (A) = Ah .
Finally, the three-term recurrence for the ph now becomes the three-term
recurrence for the Ah that defines distance-regular graphs.
Noting that pd (k) depends on the spectrum only, we see that this provides a
characterization of distance-regularity in terms of the spectrum and the number
of pairs of vertices far apart (at mutual distance d). See [115], [143], [144] and
Theorem 13.5.3 below.
11.11 Exercises
Exercise 1 Determine the spectrum of a strongly regular graph minus a vertex.
(Hint: if the strongly regular graph has characteristic polynomial p(x) = (x −
k)(x − r)f (x − s)g , then the graph obtained after removing one vertex has
characteristic polynomial ((x − k)(x − λ + µ) + µ) (x − r)f −1 (x − s)g−1 .)
Determine the spectrum of a strongly regular graph minus two (non)adjacent

vertices.
Show that the spectrum of a distance-regular graph minus a vertex does not
depend on the vertex chosen. Give an example of two nonisomorphic cospectral
graphs both obtained by removing a vertex from the same distance-regular
graph.
Chapter 12
p-ranks
Designs or graphs with the same parameters can sometimes be distinguished

by considering the p-rank of associated matrices. For example, there are three
nonisomorpic 2-(16,6,2) designs, with point-block incidence matrices of 2-rank
6, 7 and 8 respectively.
Tight bounds on the occurrence of certain configurations are sometimes ob-
tained by computing a rank in some suitable field, since p-ranks of integral
matrices may be smaller than their ranks over R.
Our first aim is to show that given the parameters (say, the real spectrum),
only finitely many primes p are of interest.
12.1 Reduction mod p

A technical difficulty is that one would like to talk about eigenvalues that are
zero or nonzero mod p for some prime p, but it is not entirely clear what that
might mean when the eigenvalues are nonintegral.
√ √ Necessarily some arbitrari-
ness will be involved. For example (5 + 2)(5 − 2) ≡ 0√mod 23 and one point
of view is that this means that 23 is not a prime in Q( 2), and one gets into
algebraic number theory. But another point of view is that if one ‘reduces mod
23’, mapping to a field of characteristic
√ 23,
√ then at least one factor must become
0. However, the sum of 5 + 2 and 5 − 2 does not become 0 upon reduction
mod 23, so not both factors become 0. Since these factors are conjugate, the
‘reduction mod 23’ cannot be defined canonically, it must involve some arbitrary
choices. We follow Isaacs [205], who follows Brauer.
Let R be the ring of algebraic integers in C, and let p be a prime. Let M
be a maximal ideal in R containing the ideal pR. Put F = R/M . Then F is a
field of characteristic p. Let r 7→ r̄ be the quotient map R → R/M = F . This
will be our ‘reduction mod p’. (It is not canonical because M is not determined
uniquely.)
Lemma 12.1.1 (Isaacs [205], 15.1) Let U = {z ∈ C | z m = 1 for some integer
m not divisible by p}. Then the quotient map R → R/M = F induces an
isomorphism of groups U → F ∗ from U onto the multiplicative group F ∗ of F .
Moreover, F is algebraically closed, and is algebraic over its prime field.
One consequence is that on integers ‘reduction mod p’ has the usual meaning:
if m is an integer not divisible by p then some power is 1 (mod p) and it follows
165
166 CHAPTER 12. P -RANKS
that m̄ 6= 0. More generally, if θ̄ = 0, then p|N (θ), where N (θ) is the norm
of θ, the product of its conjugates, up to sign the constant term of its minimal
polynomial.
12.2 The minimal polynomial

Let M be a matrix of order n over a field F . For each eigenvalue θ of M in F ,
let m(θ) be the geometric multiplicity of θ, so that rk(M − θI) = n − m(θ).
Let e(θ) be the algebraic multiplicity of the eigenvalue θ,Qso that the char-
acteristic polynomial of M factors as c(x) := det(xI − M ) = (x − θ)e(θ) c0 (x),
where c0 (x) has no roots in F . Then m(θ) ≤ e(θ).
The minimal polynomial p(x) of M is the unique monic polynomial over F
of minimal degree such that p(M ) = 0. The numbers θ ∈ F for which p(θ) = 0
are precisely the eigenvalues ofQM (in F ). By Cayley-Hamilton, p(x) divides
c(x). It follows that if p(x) = (x − θ)h(θ) p0 (x), where p0 (x) has no roots in
F , then 1 ≤ h(θ) ≤ e(θ).
In terms of the Jordan decomposition of M , m(θ) is the number of Jordan
blocks for θ, h(θ) is the size of the largest block, and e(θ) is the sum of the sizes
of all Jordan blocks for θ.
We see that n − e(θ) + h(θ) − 1 ≤ rk(M − θI) ≤ n − e(θ)/h(θ), and also that
1 ≤ rk((M − θI)i ) − rk((M − θI)i+1 ) ≤ m(θ) for 1 ≤ i ≤ h − 1.
12.3 Bounds for the p-rank

Let M be a square matrix of order n, and let rkp (M ) be its p-rank. Let R and
F be as above in §12.1. Use a suffix F or p to denote rank or multiplicity over
the field F or Fp (instead of C).
Proposition 12.3.1
Let M be an integral square matrix. Then rkp (M ) ≤ rk(M ).
Let M be a square matrix with entries in R. Then rkF (M ) ≤ rk(M ).
Proof. The rank of a matrix is the size of the largest submatrix with nonzero
determinant.
Proposition 12.3.2 Let M be an integral square matrix. Then

X
rkp (M ) ≥ {m(θ) | θ̄ 6= 0}.
Proof. LetP M have order n. P Then rkp (M ) = n P− mp (0) ≥ n − ep (0) =

n − eF (0) = {eF (t) | t 6= 0} = {e(θ) | θ̄ 6= 0} ≥ {m(θ) | θ̄ 6= 0}.
Proposition 12.3.3 Let the integral square matrix M be diagonalizable. Then

rkF (M − θ̄I) ≤ n − e(θ) for each eigenvalue θ of M .
Proof. rkF (M − θ̄I) ≤ rk(M − θI) = n − m(θ) = n − e(θ).

It follows that if θ̄ = 0 for a unique θ, then rkp (M ) = n − e(θ). We can still say
something when θ̄ = 0 for two eigenvalues θ, when one has multiplicity 1:
12.4. INTERESTING PRIMES P 167
Proposition 12.3.4 Let the integral square matrix M be diagonalizable, and

suppose that θ̄ = 0 for only two eigenvalues θ, say θ0 and θ1 , where e(θ0 ) = 1.
Let M have minimal polynomial p(x) = (x − θ0 )f (x). Then rkF (M ) = n −
e(θ1 ) − ε, where ε = 1 if f (M ) = 0 and ε = 0 otherwise.
Proof. By the above n − e(θ1 ) − 1 ≤ rkF (M ) ≤ n − e(θ1 ). By the previous

section n − eF (0) + hF (0) − 1 ≤ rkF (M ) ≤ n − eF (0)/hF (0). Since eF (0) =
e(θ1 ) + 1 we find rkF (M ) = n − e(θ1 ) − ε, where ε = 1 if hF (0) = 1 and ε = 0
otherwise. But hF (0) = 1 iff f (M ) = 0.
If M is a matrix with integral entries, then the minimal polynomial p(x) and its
factor f (x) have integral coefficients. In particular, if M is an integral symmetric
matrix with constant row sums k, and the eigenvalue k of M has multiplicity
1, thenQf (M ) = (f (k)/n)J and the condition f (M ) = 0 becomes c̄ = 0, where
c = n1 θ6=k (k − θ) is an integer.
12.4 Interesting primes p

Let A be an integral matrix of order n, and let M = A − aI for some integer a.
If θ is an eigenvalue of A, then θ − a is an eigenvalue of M .
If θ̄ = ā for no θ, then rkp (M ) = n.
If θ̄ = ā for a unique θ, then rkp (M ) = rkF (M ) = rkF (M − (θ¯ − a)I)
≤ rk(A − θI) = n − m(θ) by Proposition 12.3.1, but also rkp (M ) ≥ n − m(θ) by
Proposition 12.3.2, so that rkp (M ) = n − m(θ).
So, if the p-rank of M is interesting, if it gives information not derivable
from the spectrum of A and the value a, then at least two eigenvalues of M
become zero upon reduction mod p. But if θ − a = η − a = 0, then θ − η = 0,
and in particular p|N (θ − η), which happens for finitely many p only.
Example The unique distance-regular graph with intersection array {4, 3, 2;
√ 6 √
1, 2, 4} has 14 vertices and spectrum 4, 2 , (− 2)6 , −4 (with multiplicities
written as exponents).
Let A be the adjacency matrix of this graph, and consider the p-rank of
M = A − aI for integers a. The norms of θ − a are 4 − a, a2 − 2, −4 − a, and
if these are all nonzero mod p then the p-rank of M is 14. If p is not 2 or 7,
then at most one of these norms can be 0 mod p, and for a ≡ 4 (mod p) or
2
a ≡ −4 (mod p) the √ p-rank of M √ is 13. If a ≡ 2 (mod p) then precisely one
of the eigenvalues 2 − a and − 2 − a reduces to 0, and the p-rank of M is 8.
Finally, for p = 2 and p = 7 we need to look at the matrix M itself, and find
rk2 (A) = 6 and rk7 (A ± 3I) = 8.
12.5 Adding a multiple of J

Let A be an integral matrix of order n with row and column sums k, and consider
the rank and p-rank of M = Mb = A+bJ. Since J has rank 1, all these matrices
differ in rank by at most 1, so either all have the same rank r, or two ranks r,
r + 1 occur, and in the latter case rank r + 1 occurs whenever the row space of
M contains the vector 1.
The matrix M has row sums k + bn.
If p ∤ n, then the row space of M over Fp contains 1 when k+bn 6≡ 0 (mod p).
On the other hand, if k + bn ≡ 0 (mod p), then all rows have zero row sum
(mod p) while 1 has not, so that 1 is not in the row space over Fp . Thus, we
are in the second case, where the smaller p-rank occurs for b = −k/n only.
If p|n and p ∤ k, then all row sums are nonzero (mod p) for all b, and we
are in the former case: the rank is independent of b, and the row space over Fp
always contains 1.
Finally, if p|n and also p|k, then further inspection is required.
Example (Cf. Peeters [258]). According to [183], there are precisely ten graphs
√ 8 √
with the spectrum 71 7 (−1)7 (− 7)8 , one of which is the Klein graph, the
unique distance-regular graph with intersection array {7, 4, 1; 1, 2, 7}. It turns
out that the p-ranks of A − aI + bJ for these graphs depend on the graph only
for p = 2 ([258]). Here n = 24 and k = 7 − a + 24b.
graph rk2 (A + I) rk2 (A + I + J)
#1,2 14 14 rk3 (A − aI + bJ) rk7 (A − aI)
#3,8,9 15 14 a\b 0 1 2 a\b 0
#4,7 13 12 0 24 24 24 0 15
#5 12 12 1 16 15 16 1–5 24
#6 11 10 2 16 16 16 6 17
#10 9 8
Interesting primes (dividing the norm of the difference of two eigenvalues) are 2,
3 and 7. All p-ranks follow from the parameters except possibly rk2 (A + I + bJ),
rk3 (A − I + bJ), rk3 (A + I), rk7 (A).
The interesting 2-rank is rk2 (A + I), and inspection of the graphs involved
shows that this takes the values 9, 11, 12, 13, 14, 15 where 9 occurs only for the
Klein graph. The value of rk2 (A + I + J) follows, since a symmetric matrix with
zero diagonal has even 2-rank, and the diagonal of a symmetric matrix lies in the
F2 -space of its rows. Hence if rk2 (A+I) is even, then rk2 (A+I +J) = rk2 (A+I),
and if rk2 (A + I) is odd then rk2 (A + I + J) = rk2 (A + I) − 1.
The 3-rank of A − I + bJ is given by Proposition 12.3.4. Here f (x) =
(x + 2)((x + 1)2 − 7) and k = 6 + 24b, so that f (k)/n ≡ 0 (mod 3) is equivalent
to b ≡ 1 (mod 3).
One has rk3 (A + I) = 16 in all ten cases.
The value of rk7 (A) can be predicted: We have det(A + J) = −78 .31, so the
Smith Normal Form (§12.8) of A + J has at most 8 entries divisible by 7 and
rk7 (A + J) ≥ 16. By Proposition 12.3.3, rk7 (A + J) = 16. Since 7 ∤ n and 1 is
in the row space of A + J but not in that of A, rk7 (A) = 15.
12.6 Paley graphs

Let q be a prime power, q ≡ 1 (mod 4), and let Γ be the graph with vertex
set Fq where two vertices are adjacent whenever their difference is a nonzero
square. (Then Γ is called the Paley graph of order q.) In order to compute the
p-rank of the Paley graphs, we first need a lemma.
Pd−1 Pe−1
Lemma 12.6.1 Let p(x, y) = i=0 j=0 cij xi y j be a polynomial with coeffi-
cients in a field F . Let A, B ⊆ F , with m := |A| ≥ d and n := |B| ≥ e.
Consider the m × n matrix P = (p(a, b))a∈A, b∈B and the d × e matrix C = (cij ).
Then rkF (P ) = rkF (C).
Proof. For any integer s and subset X of F , let Z(s, X) be the |X| × s matrix
(xi )x∈X, 0≤i≤s−1 . Note that if |X| = s then this is a Vandermonde matrix and
hence invertible. We have P = Z(d, A) C Z(e, B)⊤ , so rkF (P ) ≤ rkF (C), but
P contains a submatrix Z(d, A′ ) C Z(e, B ′ ) with A′ ⊆ A, B ′ ⊆ B, |A′ | = d,
|B ′ | = e, and this submatrix has the same rank as C.
For odd prime powers q = pe , p prime, let Q be the {0, ±1}-matrix of order
q with entries Qxy = χ(y − x) (x, y ∈ Fq , χ the quadratic residue character,
χ(0) = 0).
Proposition 12.6.2 ([49]) rkp Q = ((p + 1)/2)e .
Proof. Applying the above lemma with p(x, y) = χ(y − x) = (y − x)(q−1)/2 =

P
i (q−1)/2 i (q−1)/2−i
i (−1) i xy

, we see that rkp Q equals the number of binomial
coefficients (q−1)/2
i withP0 ≤ i ≤ (q − 1)/2
P not divisible by p. Now Lucas’
Theorem says that if l = i li pi and k = i ki pi are the p-ary expansions of l
Q P
and k, then kl ≡ i klii (mod p). Since 21 (q − 1) = i 21 (p − 1)pi , this means
that for each p-ary digit of i there are (p + 1)/2 possibilities and the result
follows.
For Lucas’ Theorem, cf. MacWilliams & Sloane [236], §13.5, p. 404 (and
references given there). Note that this proof shows that each submatrix of Q of
order at least (q + 1)/2 has the same rank as Q.
The relation between Q here and the adjacency matrix A of the Paley graph
is Q = 2A + I − J. From Q2 = qI − J ≡ −J (mod p) and (2A + I)2 =
qI + (q − 1)J ≡ −J (mod p) it follows that both hQi and h2A + Ii contain 1,
so rkp (A + 21 I) = rkp (2A + I) = rkp (Q) = ((p + 1)/2)e .
12.7 Strongly regular graphs

Let Γ be a strongly regular graph with adjacency matrix A, and assume that
A has integral eigenvalues k, r, s with multiplicities 1, f, g, respectively. We
investigate the p-rank of a linear combination of A, I and J.
The following proposition shows that only the case p|(r − s) is interesting.
Proposition 12.7.1 Let M = A + bJ + cI. Then M has eigenvalues θ0 =

k + bv + c, θ1 = r + c, θ2 = s + c, with multiplicities m0 = 1, m1 = f , m2 = g,
respectively.
(i) If none of the θi vanishes (mod p), then rkp M = v.
(ii) If precisely one θi vanishes (mod p), then M has p-rank v − mi .
Put e := µ + b2 v + 2bk + b(µ − λ).
(iii) If θ0 ≡ θ1 ≡ 0 (mod p), θ2 6≡ 0 (mod p), then rkp M = g if and only if
p|e, and rkp M = g + 1 otherwise.
(iii)’ If θ0 ≡ θ2 ≡ 0 (mod p), θ1 6≡ 0 (mod p), then rkp M = f if and only if
p|e, and rkp M = f + 1 otherwise.
(iv) In particular, if k ≡ r ≡ 0 (mod p) and s 6≡ 0 (mod p), then rkp A = g.
And if k ≡ s ≡ 0 (mod p) and r 6≡ 0 (mod p), then rkp A = f .
(v) If θ1 ≡ θ2 ≡ 0 (mod p), then rkp M ≤ min(f + 1, g + 1).
Proof. Parts (i) and (ii) and (v) are immediate from Propositions 12.3.1,
12.3.2. Suppose θ0 ≡ θ1 ≡ 0 (mod p), θ2 6= 0 (mod p). Then we can apply
Proposition 12.3.4 with the two eigenvalues 0 and θ2 . Since rkp (M −θ2 I) = v−g,
and g ≤ rkp M ≤ g + 1, it follows that rkp M = g if and only if M (M − θ2 I) ≡ 0
(mod p). But using (A − rI)(A − sI) = µJ and r + s = λ − µ, we find
M (M − θ2 I) ≡ (A + bJ − rI)(A + bJ − sI) = eJ. Part (iii)’ is similar.
Thus, the only interesting case (where the structure of Γ plays a rôle) is that
where p divides both θ1 and θ2 , so that p | (r − s). In particular, only finitely
many primes are of interest. In this case we only have the upper bound (v).
Looking at the idempotents sometimes improves this bound by 1: We have
E1 = (r − s)−1 (A − sI − (k − s)v −1 J) and E2 = (s − r)−1 (A − rI − (k − r)v −1 J).
Thus, if k −s and v are divisible by the same power of p (so that (k −s)/v can be
interpreted in Fp ), then rkp (A − sI − (k − s)v −1 J) ≤ rk E1 = f , and, similarly, if
k −r and v are divisible by the same power of p then rkp (A−rI −(k −r)v −1 J) ≤
rk E2 = g.
For M = A + bJ + cI and p|(r + c), p|(s + c) we have M E1 = JE1 = 0 (over
Fp ) so that rkp hM, 1i ≤ g + 1, and hence rkp M ≤ g (and similarly rkp M ≤ f )
in case 1 ∈/ hM i.
Much more detail is given in [49] and [257].
In the table below we give for a few strongly regular graphs for each prime p
dividing r−s the p-rank of A−sI and the unique b0 such that rkp (A−sI −b0 J) =
rkp (A − sI − bJ) − 1 for all b 6= b0 , or ‘-’ in case rkp (A − sI − bJ) is independent
of b. (When p ∤ v we are in the former case, and b0 follows from the parameters.
When p|v and p ∤ µ, we are in the latter case.)
For a description of most of these graphs, see [56].
Name v k λ µ rf sg p rkp (A − sI) b0

Folded 5-cube 16 5 0 2 110 (−3)5 2 6 -
Schläfli 27 16 10 8 46 (−2)20 2 6 0
3 7 -
T (8) 28 12 6 4 47 (−2)20 2 6 0
3 8 2
3 Chang graphs 28 12 6 4 47 (−2)20 2 8 -
3 8 2
G2 (2) 36 14 4 6 221 (−4)14 2 8 -
3 14 -
Sp4 (3) 40 12 2 4 224 (−4)15 2 16 -
3 11 1
O5 (3) 40 12 2 4 224 (−4)15 2 10 -
3 15 1
Hoffman-Singleton 50 7 0 1 228 (−3)21 5 21 -
Gewirtz 56 10 0 2 235 (−4)20 2 20 -
3 20 1
M22 77 16 0 4 255 (−6)21 2 20 0
Brouwer-Haemers 81 20 1 6 260 (−7)20 3 19 -
Higman-Sims 100 22 0 6 277 (−8)22 2 22 -
5 23 -
Hall-Janko 100 36 14 12 636 (−4)63 2 36 0
5 23 -
continued...
12.8. SMITH NORMAL FORM 171
Name v k λ µ rf sg p rkp (A − sI) b0

GQ(3, 9) 112 30 2 10 290 (−10)21 2 22 -
3 20 1
001... in S(5, 8, 24) 120 42 8 18 299 (−12)20 2 20 -
7 20 5
Sp4 (5) 156 30 4 6 490 (−6)65 2 66 -
5 36 1
Sub McL 162 56 10 24 2140 (−16)21 2 20 0
3 21 -
Edges of Ho-Si 175 72 20 36 2153 (−18)21 2 20 0
5 21 -
01... in S(5, 8, 24) 176 70 18 34 2154 (−18)21 2 22 -
5 22 3
switched version 176 90 38 54 2153 (−18)22 2 22 -
of previous 5 22 3
Cameron 231 30 9 3 955 (−3)175 2 55 1
3 56 1
Berlekamp-van Lint-Seidel 243 22 1 2 4132 (−5)110 3 67 -
Delsarte 243 110 37 60 2220 (−25)22 3 22 -
S(4, 7, 23) 253 112 36 60 2230 (−26)22 2 22 0
7 23 5
McLaughlin 275 112 30 56 2252 (−28)22 2 22 0
3 22 1
5 23 -
switched version 276 140 58 84 2252 (−28)23 2 24 -
of previous plus 3 23 2
isolated point 5 24 3
G2 (4) 416 100 36 20 2065 (−4)350 2 38 -
3 65 1
Dodecads mod 1 1288 792 476 504 81035 (−36)252 2 22 0
11 230 3
Table 12.1: p-ranks of some strongly regular graphs
12.8 Smith Normal Form

The Smith Normal Form S(M ) of an integral matrix M is a diagonal matrix
S(M ) = P M Q = diag(s1 , . . . , sn ), where P and Q are integral with determinant
±1 and s1 |s2 | · · · |sn . It exists and is uniquely determined up to the signs of the
si . The si are called the elementary
divisors
or invariant factors. For example,
if M = 13 11 11 11 , then S(M ) = 10 02 00 00 .
Let hM i denote the row space of M over Z. By the fundamental theorem for
finitely generated abelian groups, the group Zn /hM i is isomorphic to a direct
sum Zs1 ⊕ · · · ⊕ Zsm ⊕ Zs for certain s1 , . . . , sm , s, where s1 | · · · |sm . Since
Zn /hM i ∼ = Zn /hS(M )i, we see that diag(s1 , . . . , sm , 0t ) is the Smith Normal
Form of M , when M has r rows and n = m + s columns, and t = min(r, n) − m.
Q Qt
If M is square then si = det S(M ) = ± det M . More generally, i=1 si is
the g.c.d. of all minors of M of order t.
The Smith Normal Form is a finer invariant than the p-rank: the p-rank is
just the number of si not divisible by p. (It follows that if M is square and
pe || det M , then rkp M ≥ n − e.)
We give some examples of graphs distinguished by Smith Normal Form or

p-rank.
Example Let A and B be the adjacency matrices of the lattice graph K4 ×
K4 and the Shrikhande graph. Then S(A) = S(B) = diag(16 , 24 , 45 , 12), but
S(A + 2I) = diag(16 , 81 , 09 ) and S(B + 2I) = diag(16 , 21 , 09 ). All have 2-rank
equal to 6.
Example An example where the p-rank suffices to distinguish, is given by the
Chang graphs, strongly regular graphs with the same parameters as the triangu-
lar graph T (8), with (v, k, λ, µ) = (28, 12, 6, 4) and spectrum 121 47 (−2)20 . If A
is the adjacency matrix of the triangular graph and B that of one of the Chang
graphs then S(A) = diag(16 , 215 , 86 , 241 ) and S(B) = diag(18 , 212 , 87 , 241 ), so
that A and B have different 2-rank.
Example Another example is given by the point graph and the line graph of
the GQ(3, 3) constructed in §8.6.2. The 2-ranks of the adjacency matrices are
10 and 16 respectively.
u u u
@
@u u
B
uB
A B
u uAB
@ @AB
u @u u @
ABu
Figure 12.1: Graphs with same Laplacian SNF (13 , 5, 15, 0)
Concerning the Smith Normal Form of the Laplacian, Grone, Merris &
Watkins [167] gave the pair of graphs in Figure 12.1 that both have S(L) =
diag(13 , 5, 15, 0). The Laplacian spectrum of the left one (which is K2 × K3 ) is
0, 2, 32 , 52 . That of the right one is 0, 0.914, 3.572, 52 , 5.514, where the three
non-integers are roots of λ3 − 10λ2 + 28λ − 18 = 0.
12.8.1 Smith Normal Form and spectrum

There is no very direct connection
between Smith Normal Form and spectrum.
3 1
For example, the matrix has eigenvalues 2 and 4, and invariant factors 1
1 3
and 8.
Proposition 12.8.1 Let M be an integral matrix of order n, with invariant

factors s1 , . . . , sn .
(i) If a is an integral eigenvalue of M , then a|sn .
(ii) If a is an integral eigenvalue of M with geometric multiplicity m, then
a|sn−m+1 .
(iii) If M is diagonalizable with distinct eigenvalues a1 , . . . , am , all integral,
then we have sn |a1 a2 · · · am .
12.8. SMITH NORMAL FORM 173
Proof. Part (i) is a special case of (ii). Part (ii) is Proposition 12.8.4 below.
For (iii) we may assume that all ai are nonzero. It suffices to show that every
element in Zn /hM P i has an order dividing a1 a2 · · · am . We show by induction on
k that if u = ui is integral and is sum of k left eigenvectors
P ui of M , with
then a1 · · · ak u ∈ hM i. Indeed, since uM =
u i M = ai u i , P ai ui ∈ hM i and
ak u − uM = (ak − ai )ui is integral and sum of at most k − 1 eigenvectors, we
find by induction that a1 · · · ak−1 (ak u − uM ) ∈ hM i, hence a1 · · · ak u ∈ hM i.
The invariant factors are determined when we know for each prime p and
each i ≥ 0 how many invariant factors are divisible by pi , and the following
proposition tells us.
Proposition 12.8.2 Let A be an integral matrix of order n, p a prime number

and i a nonnegative integer. Put Mi := Mi (A) := {x ∈ Zn | p−i Ax ∈ Zn }. Let
Mi ⊆ Fpn be the mod p reduction of Mi . Then Mi is an Fp -vectorspace, and the
number of invariant factors of A divisible by pi equals dimp Mi .
Proof. dimp Mi does not change when A is replaced by P AQ where P and Q

are integral matrices of determinant 1. So we may assume that A is already in
Smith Normal Form. Now the statement is obvious.
There is a dual statement:
Proposition 12.8.3 Let A be an integral matrix of order n, p a prime number

and i a nonnegative integer. Put Ni := Ni (A) := {p−i Ax | x ∈ Mi }. Then the
number of invariant factors of A not divisible by pi+1 equals dimp Ni .
Proof. dimp Ni does not change when A is replaced by P AQ where P and Q

are integral matrices of determinant 1. So we may assume that A is already in
Smith Normal Form. Now the statement is obvious.
Proposition 12.8.4 Let A be a square integral matrix with integral eigenvalue

a of (geometric) multiplicity m. Then the number of invariant factors of A
divisible by a is at least m.
Proof. Let W = {x ∈ Qn | Ax = ax} be the a-eigenspace of A over Q, so that

dimQ (W ) = m. By the Proposition 12.8.2 it suffices to show that dimp W = m
for all primes p, where W is the mod p reduction of W ∩ Zn . Pick a basis
x1 , . . . , xm of W consisting of m integral vectors, chosen in such a way that the
n × m matrix X that has columns xj has a (nonzero) minor of order m with the
minimum possible number of factors P p. If upon reduction mod p these Pvectors
become dependent, that is, if cj xj = 0 wherePnot all cj vanish, then cj xj
has coefficients divisible by p, so that y := p1 cj xj ∈ W ∩ Zn , and we can
replace some xj (with nonzero cj ) by y and get a matrix X ′ where the minors
have fewer factors p, contrary to assumption. So, the xi remain independent
upon reduction mod p, and dimp W = m.
Example Let q = pt for some prime p. Consider the adjacency matrix A of

the graph Γ of which the vertices are the lines of P G(3, q), where two lines are
adjacent when they are disjoint. This graph is strongly regular, with eigenvalues
k = q 4 , r = q, s = −q 2 and multiplicities 1, f = q 4 + q 2 , g = q 3 + q 2 + q,
respectively. Since det A is a power of p, all invariant factors are powers of p.
Let pi occur as invariant factor with multiplicity ei .
Claim. We have e0 + e1 + · · · + et = f and e2t + · · · + e3t = g and e4t = 1

and ei = 0 for t 4t. Moreover, e3t−i = ei for
0 ≤ i < t.
Proof.
P The total number of invariant factors is the P size of the matrix, so
e =
i i P f + g + 1. The number of factors p in det A is i iei = t(f + 2g + 4).
Hence i (i − P t)ei = t(g + 3).
Let mi := j≥i ej . By (the proof of) Proposition 12.8.4 we have m4t ≥ 1
and m2t ≥ g + 1. (The +1 follows because 1 is orthogonal to eigenvectors with
eigenvalue other than k, but has a nonzero (mod p) inner product with itself,
so that 1 6∈ W for an eigenspace W with 1 6∈ W .)
The matrix A satisfies the equation (A − rI)(A − sI) = µJ, that is, A(A +
q(q − 1)I) = q 3 I + q 3 (q − 1)J, and the right-hand side is divisible by p3t . If
x ∈ Zn and p−i (A + q(q − 1)I)x ∈ Zn , then p−i (A + q(q − 1)I)x ∈ M3t−i (A) for
0 ≤ i ≤ 3t. If 0 ≤ i < t, then p−i q(q − 1)x = 0 (mod p), so that Ni ⊆ M3t−i .
Also 1 ∈ M3t−i , while 1 6∈ Ni because 1⊤ p−i Ax = p4t−i Ax reduces to 0 (mod p)
for integral x, unlike 1⊤ 1. By Proposition 12.8.3 we find m3t−i ≥ e0 +· · ·+ei +1
(0 ≤ i < t). P P
Adding the inequalities − 0≤i≤h ei + i≥3t−h ei ≥ 1 (0 ≤ h < t), and
P P
t i≥2t ei ≥ t(g + 1) and t i≥4t ei ≥ t yields
X X X X
(i − t)ei + (i − t)ei + 2t ei + 3t ei ≥ t(g + 3)
0≤i<t 2t≤i≤3t 3t+1≤i<4t i≥4t
P
and equality must hold everywhere since i (i − t)ei = t(g + 3).
Note that our conclusion also holds for any strongly regular graph with the
same parameters as this graph on the lines of P G(3, q).
In the particular case q = p, the invariant factors are 1, p, p2 , p3 , p4 with
multiplicities e, f − e, g − e, e, 1, respectively, where e = 13 p(2p2 + 1) in the case
of the lines of P G(3, p) (cf. [135]). Indeed, the number e of invariant factors
not divisible by p is the p-rank of A, determined in Sin [284].
For p = 2, there are 3854 strongly regular graphs with parameters (35,16,6,8)
([244]), and the 2-ranks occurring are 6, 8, 10, 12, 14 (with frequencies 1, 3, 44,
574, 3232, respectively)—they must be even because A is alternating (mod 2).
The invariant factors of the disjointness graph of the lines of P G(3, 4) are
136 216 4220 1632 3216 6436 2561 , with multiplicities written as exponents.
One can generalize the above observations, and show for example that if p is
a prime, and A is the adjacency matrix of a strongly regular graph, and pa ||k,
pb ||r, pc ||s, where a ≥ b + c and p ∤ v, and A has ei invariant factors sj with
pi ||sj , then ei = 0 for min(b, c) a.
Moreover, eb+c−i = ei for 0 ≤ i < min(b, c).
Chapter 13
Spectral characterizations
In this chapter, we consider the question to what extent graphs are determined
by their spectrum. First we give several constructions of families of cospec-
tral graphs, and then give cases in which it has been shown that the graph is
determined by its spectrum.
Let us abbreviate ‘determined by the spectrum’ to DS. Here, of course,
‘spectrum’ (and DS) depends on the type of adjacency matrix. If the matrix is
not specified, we mean the ordinary adjacency matrix.
Large parts of this chapter were taken from Van Dam & Haemers [117, 118].
13.1 Generalized adjacency matrices

Let A = AΓ be the adjacency matrix of a graph Γ. The choice of 0, 1, 0 in A to
represent equality, adjacency and non-adjacency was rather arbitrary, and one
can more generally consider a matrix xI + yA + z(J − I − A) that uses x, y, z
instead. Any such matrix, with y 6= z, is called a generalized adjacency matrix
of Γ. The spectrum of any such matrix is obtained by scaling and shifting from
that of a matrix of the form A + yJ, so for matters of cospectrality we can
restrict ourselves to this case.
Call two graphs Γ and ∆ y-cospectral (for some real y) when AΓ − yJ and
A∆ −yJ have the same spectrum. Then 0-cospectral is what we called cospectral,
and 12 -cospectral is Seidel-cospectral, and 1-cospectrality is cospectrality for the
complementary graphs. Call two graphs just y-cospectral when they are y-
cospectral but not z-cospectral for any z 6= y.
The graphs K1,4 and K1 + C4 are just 0-cospectral. The graphs 2K3 and
2K1 + K4 are just 31 -cospectral. The graphs K1 + C6 and Ê6 (cf. §1.3.7) are
y-cospectral for all y.
Proposition 13.1.1
(i) (Johnson & Newman [208]) If two graphs are y-cospectral for two distinct
values of y, then for all y.
(ii) (Van Dam, Haemers & Koolen [119]) If two graphs are y-cospectral for an
irrational value of y, then for all y.
175
176 CHAPTER 13. SPECTRAL CHARACTERIZATIONS
Proof. Define p(x, y) = det(AΓ − xI − yJ). Thus for fixed y, p(x, y) is the
characteristic polynomial of AΓ − yJ. Since J has rank 1, the degree in y of
p(x, y) is 1 (this follows from Gaussian elimination in AΓ − xI − yJ), so there
exist integers a0 , . . . , an and b0 , . . . , bn such that
n
X
p(x, y) = (ai + bi y)xi .
i=0
Suppose Γ and Γ′ are y-cospectral for some y = y0 but not for all y. Then
the corresponding polynomials p(x, y) and p′ (x, y) are not identical, whilst
p(x, y0 ) = p′ (x, y0 ). This implies that ai + bi y0 = a′i + b′i y0 with bi 6= b′i for
some i. So y0 = (a′i − ai )/(bi − b′i ) is unique and rational.
Van Dam, Haemers & Koolen [119] show that there is a pair of nonisomorphic
just y-cospectral graphs if and only if y is rational.
Values of y other than 0, 12 , 1 occur naturally when studying subgraphs of
strongly regular graphs.
Proposition 13.1.2 Let Γ be strongly regular with vertex set X of size n, and
let θ be an eigenvalue other than the valency k. Let y = (k − θ)/n. Then for
each subset S of X, the spectrum of Γ and the y-spectrum of the graph induced
on S determines the y-spectrum of the graph induced on X \ S.
Proof. Since A − yJ has only two eigenvalues, this follows immediately from
Lemma 2.11.1.
This can be used to produce cospectral pairs. For example, let Γ be the Petersen
graph, and let S induce a 3-coclique. Then the y-spectrum of the graph induced
on X \ S is determined by that on S, and does not depend on the coclique
chosen. Since θ can take two values, the graphs induced on the complement of
a 3-coclique (Ê6 and K1 + C6 ) are y-cospectral for all y.
13.2 Constructing cospectral graphs

Many constructions of cospectral graphs are known. Most constructions from
before 1988 can be found in [106, §6.1] and [105, §1.3]; see also [154, §4.6]. More
recent constructions of cospectral graphs are presented by Seress [280], who gives
an infinite family of cospectral 8-regular graphs. Graphs cospectral to distance-
regular graphs can be found in [48], [117], [183], and in §13.2.2. Notice that
the mentioned graphs are regular, so they are cospectral with respect to any
generalized adjacency matrix, which in this case includes the Laplacian matrix.
There exist many more papers on cospectral graphs. On regular, as well
as non-regular graphs, and with respect to the Laplacian matrix as well as the
adjacency matrix. We mention [41], [148], [187], [233], [245] and [264], but don’t
claim to be complete.
Here we discuss four construction methods for cospectral graphs. One used
by Schwenk to construct cospectral trees, one from incidence geometry to con-
struct graphs cospectral with distance-regular graphs, one presented by Godsil
and McKay, which seems to be the most productive one, and finally one due to
Sunada.
13.2. CONSTRUCTING COSPECTRAL GRAPHS 177
13.2.1 Trees
Let Γ and ∆ be two graphs, with vertices x and y, respectively. Schwenk [267]
examined the spectrum of what he called the coalescence of these graphs at x
and y, namely, the result Γ +x,y ∆ of identifying x and y in the disjoint union
Γ + ∆. He proved the following (see also [106, p.159] and [154, p.65]).
Lemma 13.2.1 Let Γ and Γ′ be cospectral graphs and let x and x′ be vertices
of Γ and Γ′ respectively. Suppose that Γ − x (that is the subgraph of Γ obtained
by deleting x) and Γ′ − x′ are cospectral too. Let ∆ be an arbitrary graph with
a fixed vertex y. Then Γ +x,y ∆ is cospectral with Γ′ +x′ ,y ∆.
Proof. Let z be the vertex of Z := Γ +x,y ∆ that is the result of identifying x

and y. A directed cycle in Z cannot meet both Γ − x and ∆ − y. By §1.2.1 the
characteristic polynomial p(t) of Z can be expressed in the numbers of unions
of directed cycles with given number of vertices and of components. We find
p(t) = pΓ−x (t)p∆ (t) + pΓ (t)p∆−y (t) − tpΓ−x (t)p∆−y (t).
For example, let Γ = Γ′ be as given below, then Γ − x and Γ − x′ are
cospectral, because they are isomorphic.
u u
x x′
u u u u u u u u u
Suppose ∆ = P3 and let y be the vertex of degree 2. Then Lemma 13.2.1 shows
that the graphs in Figure 13.1 are cospectral.
u u u u u u
u u u u u u u u u u u u u u u u u u
u u
Figure 13.1: Cospectral trees
It is clear that Schwenk’s method is very suitable for constructing cospectral
trees. In fact, the lemma above enabled him to prove his famous theorem:
Theorem 13.2.2 With respect to the adjacency matrix, almost all trees are
non-DS.
After Schwenk’s result, trees were proved to be almost always non-DS with
respect to all kinds of matrices. Godsil and McKay [155] proved that almost
all trees are non-DS with respect to the adjacency matrix of the complement
A, while McKay [243] proved it for the Laplacian matrix L and for the distance
matrix D.
13.2.2 Partial linear spaces

A partial linear space consists of a (finite) set of points P, and a collection L
of subsets of P called lines, such that two lines intersect in at most one point
(and consequently, two points are on at most one line). Let (P, L) be such a
partial linear space and assume that each line has exactly q points, and each
point is on q lines. Then clearly |P| = |L|. Let N be the point-line incidence
matrix of (P, L). Then N N ⊤− qI and N ⊤N − qI both are the adjacency matrix
of a graph, called the point graph (also known as collinearity graph) and line
graph of (P, L), respectively. These graphs are cospectral, since N N ⊤ and N ⊤N
are. But in many examples they are non-isomorphic. An example was given in
§11.3.1.
13.2.3 GM switching
Seidel switching was discussed above in §1.8.2. No graph with more than one
vertex is DS for the Seidel adjacency matrix. In some cases Seidel switching
also leads to cospectral graphs for the adjacency spectrum, for example when
graph and switched graph are regular of the same degree.
Godsil and McKay [156] consider a different kind of switching and give con-
ditions under which the adjacency spectrum is unchanged by this operation.
We will refer to their method as GM switching. (See also §1.8.3.) Though GM
switching has been invented to make cospectral graphs with respect to the ad-
jacency matrix, the idea also works for the Laplacian and the signless Laplacian
matrix, as will be clear from the following formulation.
Theorem 13.2.3 Let N be a (0, 1)-matrix of size b × c (say) whose column

e to be the matrix obtained from N by replacing
sums are 0, b or b/2. Define N
each column v with b/2 ones by its complement 1 − v. Let B be a symmetric
b × b matrix with constant row (and column) sums, and let C be a symmetric
c × c matrix. Put
" #
B N B Ne
M= and M f=
N⊤ C Ne⊤ C .
f are cospectral.
Then M and M
2
J − Ib 0 f.
Proof. Define Q = b . Then Q−1 = Q and QM Q−1 = M
0 Ic
The matrix partition used in [156] (and in §1.8.3) is more general than the
one presented here. But this simplified version suffices for our purposes: to show
that GM switching produces many cospectral graphs.
If M and M f are adjacency matrices of graphs then GM switching gives
cospectral graphs with cospectral complements and hence, by the result of John-
son & Newman quoted in §13.1, it produces cospectral graphs with respect to
any generalized adjacency matrix.
If one wants to apply GM switching to the Laplacian matrix L of a graph Γ,
take M = −L and let B and C (also) denote the sets of vertices indexing the
rows and columns of the matrices B and C, respectively. The requirement that
the matrix B has constant row sums means that N has constant row sums, that
is, the vertices of B all have the same number of neighbors in C.
For the signless Laplacian matrix, take M = Q. Now all vertices in B must
have the same number of neighbors in C, and, in addition, the subgraph of Γ
induced by B must be regular.
When Seidel switching preserves the valency of a graph, it is a special case
of GM switching, where all columns of N have b/2 ones. So the above theorem
also gives sufficient conditions for Seidel switching to produce cospectral graphs
with respect to the adjacency matrix A and the Laplacian matrix L.
13.3. ENUMERATION 179
If b = 2, GM switching just interchanges the two vertices of B, and we call

it trivial. But if b ≥ 4, GM switching almost always produces non-isomorphic
graphs. In Figures 13.2 and 13.3 we have two examples of pairs of cospectral
H P
H u u
@P
u u u u u u
HH HP @
P @
H @H HP@P
P @
u u u H Hu u @u H

H@u PP @u
Figure 13.2: Two graphs cospectral w.r.t. any generalized adjacency matrix
P
u
P
u u u H
u H u
u u
PP @HH @H
H
@

PP @ H
@
PP HH
H @
u u u Pu @ u

u H@u H @u
Figure 13.3: Two graphs cospectral w.r.t. the Laplacian matrix

graphs produced by GM switching. In both cases b = c = 4 and the upper
vertices correspond to B and the lower vertices to C. In the example of Figure
13.2, B induces a regular subgraph and so the graphs are cospectral with respect
to every generalized adjacency matrix. In the example of Figure 13.3 all vertices
of B have the same number of neighbors in C, so the graphs are cospectral with
respect to the Laplacian matrix L.
13.2.4 Sunada’s method

As a corollary of the discussion in §6.3 we have:
Proposition 13.2.4 Let Γ be a finite graph, and G a group of automorphisms.

If H1 and H2 are subgroups of G such that Γ is a cover of Γ/Hi (i = 1, 2) and
such that each conjugacy class of G meets H1 and H2 in the same number of
elements, then the quotients Γ/Hi (i = 1, 2) have the same spectrum and the
same Laplacian spectrum.
Sunada [289] did this for manifolds, and the special case of graphs was
discussed in [187]. See also [41].
Proof. The condition given just means that the induced characters 1G
Hi (i =
1, 2) are the same. Now apply Lemma 6.3.1 with M = A and M = L.
Brooks [41] shows a converse: any pair of regular connected cospectral graphs
arises from this construction.
13.3 Enumeration
13.3.1 Lower bounds
GM switching gives lower bounds for the number of pairs of cospectral graphs
with respect to several types of matrices.
Let Γ be a graph on n−1 vertices and fix a set X of three vertices. There is a
unique way to extend Γ by one vertex x to a graph Γ′ , such that X ∪{x} induces
a regular graph in Γ′ and that every other vertex in Γ′ has an even number of
neighbors in X ∪ {x}. Thus the adjacency matrix of Γ′ admits the structure
of Theorem 13.2.3, where B corresponds to X∪ {x}. This implies that from a
graph Γ on n − 1 vertices one can make n−1 3 graphs with a cospectral mate
on n vertices (with respect to any generalized adjacency matrix) and every such
n-vertex graph can be obtained in four ways from a graph on n − 1 vertices. Of
course some of these graphs may be isomorphic, but the probability of such a
coincidence tends to zero as n → ∞ (see [185] for details). So, if gn denotes the
number of non-isomorphic graphs on n vertices, then:
Theorem 13.3.1 The number of graphs on n vertices which are non-DS with
respect to any generalized adjacency matrix is at least
1
( 24 − o(1)) n3 gn−1 .
The fraction of graphs with

√ the required condition with b = 4 for the Laplacian
matrix is roughly 2−n n n. This leads to the following lower bound (again see
[185] for details):
Theorem 13.3.2 The number of non-DS graphs on n vertices with respect to

the Laplacian matrix is at least
√
rn ngn−1 ,
for some constant r > 0.
In fact, a lower bound like the one in Theorem 13.3.2 can be obtained for any
matrix of the form A + αD, including the signless Laplacian matrix Q.
13.3.2 Computer results

The mentioned papers [155] and [156] by Godsil and McKay also give interesting
computer results for cospectral graphs. In [156] all graphs up to 9 vertices are
generated and checked on cospectrality. This enumeration has been extended
to 11 vertices by Haemers & Spence [185], and cospectrality was tested with
respect to the adjacency matrix A, the set of generalized adjacency matrices
(A & A), the Laplacian matrix L, and the signless Laplacian matrix Q. The
results are in Table 13.1, where we give the fractions of non-DS graphs for each
of the four cases. The last three columns give the fractions of graphs for which
GM switching gives cospectral non-isomorphic graphs with respect to A, L and
Q, respectively. So column GM-A gives a lower bound for column A & A (and,
of course, for column A), column GM-L is a lower bound for column L, and
column GM-Q is a lower bound for column Q.
Notice that for n ≤ 4 there are no cospectral graphs with respect to A or to L,
but there is one such pair with respect to Q, namely K1,3 and K1 + K3 . For
n = 5 there is just one pair with respect to A. This is of course the Saltire pair
(K1,4 and K1 + C4 ).
An interesting result from the table is that the fraction of non-DS graphs is
nondecreasing for small n, but starts to decrease at n = 10 for A, at n = 9 for
L, and at n = 6 for Q. Especially for the Laplacian and the signless Laplacian
matrix, these data suggest that the fraction of DS graphs might tend to 1 as
13.4. DS GRAPHS 181
n # graphs A A&A L Q GM-A GM-L GM-Q

2 2 0 0 0 0 0 0 0
3 4 0 0 0 0 0 0 0
4 11 0 0 0 0.182 0 0 0
5 34 0.059 0 0 0.118 0 0 0
6 156 0.064 0 0.026 0.103 0 0 0
7 1044 0.105 0.038 0.125 0.098 0.038 0.069 0
8 12346 0.139 0.094 0.143 0.097 0.085 0.088 0
9 274668 0.186 0.160 0.155 0.069 0.139 0.110 0
10 12005168 0.213 0.201 0.118 0.053 0.171 0.080 0.001
11 1018997864 0.211 0.208 0.090 0.038 0.174 0.060 0.001
12 165091172592 0.188 0.060 0.027
Table 13.1: Fractions of non-DS graphs
n → ∞. In addition, the table shows that the majority of non-DS graphs with
respect to A & A and L comes from GM switching (at least for n ≥ 7). If this
tendency continues, the lower bounds given in Theorems 13.3.1 and 13.3.2 will
be asymptotically tight (with maybe another constant) and almost all graphs
will be DS for all three cases. Indeed, the fraction of graphs that admit a non-
trivial GM switching tends to zero as n tends to infinity, and the partitions with
b = 4 account for most of these switchings (see also [156]). For data for n = 12,
see [60] and [287].
13.4 DS graphs
In Section 13.2 we saw that many constructions for non-DS graphs are known,
and in the previous section we remarked that it seems more likely that almost
all graphs are DS, than that almost all graphs are non-DS. Yet much less is
known about DS graphs than about non-DS graphs. For example, we do not
know of a satisfying counterpart to the lower bounds for non-DS graphs given
in §13.3.1. The reason is that it is not easy to prove that a given graph is DS.
Below we discuss the graphs known to be DS. The approach is via structural
properties of a graph that follow from the spectrum. So let us start with a short
survey of such properties.
13.4.1 Spectrum and structure

Let us first investigate for which matrices one can see from the spectrum whether
the graph is regular.
Proposition 13.4.1 Let D denote the diagonal matrix of degrees. If a regular

graph is cospectral with a non-regular one with respect to the matrix R = A +
βJ + γD + δI, then γ = 0 and −1 < β < 0.
Proof. W.l.o.g. δ = 0. Let Γ be a graph with the given spectrum, and suppose
P di (1 ≤ i ≤ n).
that Γ has n vertices and vertex degrees
First suppose that γ 6= 0. Then i di is determined by Ptr(R) and hence
P by
the spectrum of R. Since tr(R2 ) = β 2 n2 + (1 + 2β + 2βγ) i di + γ 2 i d2i , it
P 2
follows that also
P i di is P
determined by the spectrum. Now Cauchy’s inequality
states that ( i di )2 ≤ n i d2i with equality if and only if d1 = . . . = dn . This
shows that regularity of the graph can be seen from the spectrum of R. P
Now suppose γ = 0 and β 6= −1/2. By considering tr(R2 ) we see that i di
is determined byPthe spectrum of R. The matrix R = A + βJ has average row
sum r = βn + i di /n determined by its spectrum. Let R have eigenvalues
θ1 ≥ ... ≥ θn . By interlacing, θ1 ≥ r ≥ θn , and equality on either side implies
that R has constant row sums, and Γ is regular. On the other hand, if β ≥ 0
(resp. β ≤ −1), then R (resp. −R) is a nonnegative matrix, hence if Γ is
regular, then 1 is an eigenvector for eigenvalue r = θ1 (resp. r = −θn ). Thus
also here regularity of the graph can be seen from the spectrum.
It remains to see whether one can see from the spectrum of A − yJ (with
0 < y < 1) whether the graph is regular. For y = 12 the answer is clearly no:
The Seidel adjacency matrix is S = J − I − 2A, and for S a regular graph can
be cospectral with a non-regular one (e.g. K3 and K1 + K2 ), or with another
regular one with different valency (e.g. 4K1 and C4 ). Chesnokov & Haemers
[82] constructed pairs of y-cospectral graphs where one is regular and the other
not for all rational y, 0 < y < 1. Finally, if y is irrational, then one can deduce
regularity from the spectrum of A − yJ by Proposition 13.1.1(ii).
Corollary 13.4.2 For regular graphs, being DS (or not DS) is equivalent for
the adjacency matrix, the adjacency matrix of the complement, the Laplacian,
and the signless Laplacian matrix.
Proof. For each of these matrices the above proposition says that regularity
can be recognized. It remains to find the valency k. For A, A, Q, the largest
eigenvalue is k, n − 1 − k, 2k, respectively. For L, the trace is nk.
Lemma 13.4.3 For the adjacency matrix, the Laplacian matrix and the sign-
less Laplacian matrix of a graph Γ, the following can be deduced from the spec-
trum.
(i) The number of vertices.

(ii) The number of edges.
(iii) Whether Γ is regular.
(iv) Whether Γ is regular with any fixed girth.
For the adjacency matrix the following follows from the spectrum.
(v) The number of closed walks of any fixed length.
(vi) Whether Γ is bipartite.
For the Laplacian matrix the following follows from the spectrum.
(vii) The number of components.
(viii) The number of spanning trees.
Proof. Part (i) is clear. For L and Q the number of edges is twice the trace of
the matrix, while parts (ii) and (v) for A were shown in Proposition 1.3.1. Part
(vi) follows from (v), since Γ is bipartite if and only if Γ has no closed walks
of odd length. Part (iii) follows from Proposition 13.4.1, and (iv) follows from
(iii) and the fact that in a regular graph the number of closed walks of length
less than the girth depends on the degree only. Parts (vii) and (viii) follow from
Propositions 1.3.7 and 1.3.4.
13.4. DS GRAPHS 183
H HH
u u

JHH
H

H
u
u Hu
Ju
H
u
u u Hu
HH HHJ
Hu H
Ju
Figure 13.4: Two graphs cospectral √

w.r.t. the Laplacian
√ matrix
(Laplacian spectrum: 0, 3 − 5, 2, 3, 3, 3 + 5)
The Saltire pair shows that (vii) and (viii) do not hold for the adjacency
matrix. The two graphs of Figure 13.4 have cospectral Laplacian matrices.
They illustrate that (v) and (vi) do not follow from the Laplacian spectrum.
The graphs K1 + K3 and K1,3 show that (v)–(viii) are false for the signless
Laplacian matrix.
13.4.2 Some DS graphs

Lemma 13.4.3 immediately leads to some DS graphs.
Proposition 13.4.4 The graphs Kn and Km,m and Cn and their complements
are DS for any matrix R = A + βJ + γD + δI for which regularity follows from
the spectrum of R. In particular this holds for the matrices A, A, L and R.
Proof. Since these graphs are regular, we only need to show that they are
DS with respect to the adjacency matrix. A graph cospectral with Kn has n
vertices and n(n − 1)/2 edges and therefore equals Kn . A graph cospectral with
Km,m is regular and bipartite with 2m vertices and m2 edges, so it is isomorphic
to Km,m . A graph cospectral with Cn is 2-regular with girth n, so it equals Cn .
Proposition 13.4.5 The disjoint union of k complete graphs, Km1 +. . .+Kmk ,

is DS with respect to the adjacency matrix.
Proof. The spectrum of the adjacency matrix A of any graph cospectral with
Km1 + . . . + Kmk equals {[m1 − 1]1 , . . . , [mk − 1]1 , [−1]n−k }, where n = m1 +
. . . + mk . This implies that A + I is positive semi-definite of rank k, and hence
A + I is the matrix of inner products of n vectors in Rk . All these vectors are
unit vectors, and the inner products are 1 or 0. So two such vectors coincide or
are orthogonal. This clearly implies that the vertices can be ordered in such a
way that A + I is a block diagonal matrix with all-ones diagonal blocks. The
sizes of these blocks are non-zero eigenvalues of A + I.
In general, the disjoint union of complete graphs is not DS with respect to A
and L. The Saltire pair shows that K1 + K4 is not DS for A, and K5 + 5K2
is not DS for L, because it is cospectral with the Petersen graph extended by
five isolated vertices (both graphs have Laplacian spectrum [0]6 [2]5 [5]4 ). Note
that the above proposition also shows that a complete multipartite graph is DS
with respect to A.
Proposition 13.4.6 The path with n vertices is determined by the spectrum of

its adjacency matrix. More generally, each connected graph with largest eigen-
value less than 2 is determined by its spectrum.
Proof. Let Γ be connected with n vertices and have largest eigenvalue less
than 2, and let the graph ∆ be cospectral. Then ∆ does not contain a cycle,
and has n − 1 edges, so is a tree. By Theorem 3.1.3 (and following remarks)
we find that ∆ is one of An = Pn , Dn , E6 , E7 , E8 , and has largest eigenvalue
2 cos πh , where h is the Coxeter number. Now ∆ is determined by n and h, that
is, by its number of vertices and its largest eigenvalue.
In fact, Pn is also DS with respect to A, L, and Q. The result for A, however, is
nontrivial and the subject of [134]. The hypothesis ‘connected’ here is needed,
but we can describe precisely which pairs of graphs with largest eigenvalue less
than 2 are cospectral.
Proposition 13.4.7 (i) Dn+2 + Pn is cospectral with P2n+1 + P1 for n ≥ 2.

(ii) D7 + P2 is cospectral with E6 + P3 .
(iii) D10 + P2 is cospectral with E7 + P5 .
(iv) D16 + P4 + P2 is cospectral with E8 + P9 + P5 .
(v) If two graphs Γ and ∆ with largest eigenvalue less than 2 are cospectral,
then there exist integers a, b, c such that ∆ + aP4 + bP2 + cP1 arises from Γ +
aP4 + bP2 + cP1 by (possibly repeatedly) replacing some connected components
by some others cospectral with the replaced ones according to (i)–(iv).
For example, P11 + P2 + P1 is cospectral with E6 + P5 + P3 , and P17 + P2 + P1

is cospectral with E7 + P8 + P5 , and P29 + P4 + P2 + P1 is cospectral with
E8 + P14 + P9 + P5 , and E6 + D10 + P7 is cospectral with E7 + D5 + P11 , and
E7 + D4 is cospectral with D10 + P1 , and E8 + D6 + D4 is cospectral with
D16 + 2P1 .
It follows that Pn1 + . . . + Pnk (with ni > 1 for all i) and Dn1 + . . . + Dnk
(with ni > 3 for all i) are DS.
We do not know whether Pn1 + . . . + Pnk is DS with respect to A. But it is easy
to show that this graph is DS for L and for Q.
Proposition 13.4.8 The union of k disjoint paths, Pn1 + . . . + Pnk each having
at least one edge, is DS with respect to the Laplacian matrix L and the signless
Laplacian matrix Q.
Proof. The Laplacian and the signless Laplacian eigenvalues of Pn are 2 +

2 cos πi
n , i = 1, . . . , n (see §1.4.4). Since Pn is bipartite, the signless Laplacian
eigenvalues are the same (see Proposition 1.3.10).
Suppose Γ is a graph cospectral with Pn1 + . . . + Pnk with respect to L.
Then all eigenvalues of L are less than 4. Lemma 13.4.3 implies that Γ has k
components and n1 + . . . + nk − k edges, so Γ is a forest. Let L′ be the Laplacian
matrix of K1,3 . The spectrum of L′ equals [0]1 [1]2 [4]1 . If degree 3 (or more)
occurs in Γ then L′ + D is a principal submatrix of L for some diagonal matrix
D with nonnegative entries. But then L′ + D has largest eigenvalue at least 4,
a contradiction. So the degrees in Γ are at most two and hence Γ is the disjoint
union of paths. The length m (say) of the longest path follows from the largest
eigenvalue. Then the other lengths follow recursively by deleting Pm from the
graph and the eigenvalues of Pm from the spectrum.
For a graph Γ′ cospectral with Pn1 + . . . + Pnk with respect to Q, the first
step is to see that Γ′ is a forest. But a circuit in Γ′ gives a submatrix L′ in Q
with all row sums at least 4. So L′ has an eigenvalue at least 4, a contradiction
13.4. DS GRAPHS 185
(by Corollary 2.5.2), and it follows that Γ′ is a forest and hence bipartite. Since
for bipartite graphs L and Q have the same spectrum, Γ′ is also cospectral with
Pn1 + . . . + Pnk with respect to L, and we are done.
The above two propositions show that for A, A, L, and Q the number of DS
graphs on n vertices is bounded
√ below by the number of partitions of n, which
is asymptotically equal to 2α n for some constant α. This is clearly a very poor
lower bound, but we know of no better one.
In the above we saw that the disjoint union of some DS graphs is not neces-
sarily DS. One might wonder whether the disjoint union of regular DS graphs
with the same degree is always DS. The disjoint union of cycles is DS, as can
be shown by an argument similar to that in the proof of Proposition 13.4.8.
Also the disjoint union of some copies of a strongly regular DS graph is DS. In
general we expect a negative answer, however.
13.4.3 Line graphs

The smallest adjacency eigenvalue of a line graph is at least −2 (see §1.4.5).
Other graphs with least adjacency eigenvalue −2 are the cocktailparty graphs
(mK2 , the complement of m disjoint edges) and the so-called generalized line
graphs, which are common generalizations of line graphs and cocktailparty
graphs (see [105, Ch.1]). We will not need the definition of a generalized line
graph, but only use the fact that if a generalized line graph is regular, it is a
line graph or a cocktailparty graph. Graphs with least eigenvalue −2 have been
characterised by Cameron, Goethals, Seidel and Shult [77] (cf. §7.4). They
prove that such a graph is a generalized line graph or is in a finite list of excep-
tions that comes from root systems. Graphs in this list are called exceptional
graphs. A consequence of the above characterisation is the following result of
Cvetković & Doob [104, Thm.5.1] (see also [105, Thm.1.8]).
Theorem 13.4.9 Suppose a regular graph ∆ has the adjacency spectrum of the
line graph L(Γ) of a connected graph Γ. Suppose Γ is not one of the fifteen
regular 3-connected graphs on 8 vertices, or K3,6 , or the semiregular bipartite
graph with 9 vertices and 12 edges. Then ∆ is the line graph L(Γ′ ) of a graph
Γ′ .
It does not follow that the line graph of a connected regular DS graph, which
is not one of the mentioned exceptions, is DS itself. The reason is that it can
happen that two non-cospectral graphs Γ and Γ′ have cospectral line graphs.
For example, both L(K6 ) and K6,10 have a line graph with spectrum 141 85
49 −245 , and both L(Petersen) and the incidence graph of the 2-(6, 3, 2) design
have a line graph with spectrum 61 45 14 05 −215 . The following lemma gives
necessary conditions for this phenomenon (cf. [69, Thm.1.7]).
Lemma 13.4.10 Let Γ be a k-regular connected graph on n vertices and let Γ′

be a connected graph such that L(Γ) is cospectral with L(Γ′ ). Then either Γ′ is
cospectral with Γ, or Γ′ is a semiregular bipartite graph with n + 1 vertices and
nk/2 edges, where (n, k) = (b2 − 1, ab) for integers a and b with a ≤ 21 b.
Proof. Suppose that Γ has m edges. Then L(Γ) has m vertices.

If N is the point-edge incidence matrix of Γ, then N N ⊤ is the signless
Laplacian of Γ, and N N ⊤ − kI is the adjacency matrix of Γ, and N ⊤ N − 2I is
the adjacency matrix of L(Γ). Since Γ is connected, the matrix N has eigenvalue
0 with multiplicity 1 if Γ is bipartite, and does not have eigenvalue 0 otherwise.
Consequently, L(Γ) has eigenvalue −2 with multiplicity m−n+1 if Γ is bipartite,
and with multiplicity m − n otherwise. If η 6= 0, then the multiplicity of η − 2
as eigenvalue of L(Γ) equals the multiplicity of η − k as eigenvalue of Γ.
We see that for a regular connected graph Γ, the spectrum of L(Γ) determines
that of Γ (since L(G) is regular of valency 2k − 2 and n is determined by
m = 12 nk).
Since L(Γ′ ) is cospectral with L(Γ), also Γ′ has m edges. L(Γ′ ) is regular and
hence Γ′ is regular or semiregular bipartite. Suppose that Γ′ is not cospectral
with Γ. Then Γ′ is semiregular bipartite with parameters (n1 , n2 , k1 , k2 ) (say),
and
m = 12 nk = n1 k1 = n2 k2 .
Since the signless Laplacian matrices Q and Q′ of Γ and Γ′ have the same
non-zero eigenvalues, their largest eigenvalues are equal:
2k = k1 + k2 .
If n = n1 + n2 then k1 = k2 , contradiction. So
n = n1 + n2 − 1.
Write k1 = k − a and k2 = k + a, then nk = n1 k1 + n2 k2 yields
k = (n1 − n2 )a.
Now n1 k1 = n2 k2 gives
(n1 − n2 )2 = n1 + n2 .
Put b = n1 − n2 , then (n, k) = (b2 − 1, ab). Since 2ab = k1 + k2 ≤ n2 + n1 = b2 ,
it follows that a ≤ 12 b.
Now the following can be concluded from Theorem 13.4.9 and Lemma 13.4.10.
Theorem 13.4.11 Suppose Γ is a connected regular DS graph, which is not a

3-connected graph with 8 vertices or a regular graph with b2 − 1 vertices and
degree ab for some integers a and b, with a ≤ 21 b. Then also the line graph L(Γ)
of Γ is DS.
Bussemaker, Cvetković, and Seidel [69] determined all connected regular excep-
tional graphs (see also [110]). There are exactly 187 such graphs, of which 32
are DS. This leads to the following characterisation.
Theorem 13.4.12 Suppose Γ is a connected regular DS graph with all its ad-
jacency eigenvalues at least −2, then one of the following occurs.
(i) Γ is the line graph of a connected regular DS graph.

(ii) Γ is the line graph of a connected semiregular bipartite graph, which is DS
with respect to the signless Laplacian matrix.
(iii) Γ is a cocktailparty graph.
(iv) Γ is one of the 32 connected regular exceptional DS graphs.
13.5. DISTANCE-REGULAR GRAPHS 187
Proof. Suppose Γ is not an exceptional graph or a cocktailparty graph. Then

Γ is the line graph of a connected graph ∆, say. Whitney [305] has proved
that ∆ is uniquely determined from Γ, unless Γ = K3 . If this is the case then
Γ = L(K3 ) = L(K1,3 ), so (i) holds. Suppose ∆′ is cospectral with ∆ with respect
to the signless Laplacian Q. Then Γ and L(∆′ ) are cospectral with respect to
the adjacency matrix, so Γ = L(∆′ ) (since Γ is DS). Hence ∆ = ∆′ . Because Γ
is regular, ∆ must be regular, or semiregular bipartite. If ∆ is regular, DS with
respect to Q is the same as DS.
All four cases from Theorem 13.4.12 do occur. For (i) and (iv) this is obvious,
and (iii) occurs because the cocktailparty graphs mK2 are DS (since they are
regular and A-cospectral by Proposition 13.4.5). Examples for Case (ii) are the
complete graphs Kn = L(K1,n ) with n 6= 3. Thus the fact that Kn is DS implies
that K1,n is DS with respect to Q if n 6= 3.
13.5 Distance-regular graphs

All regular DS graphs constructed so far have the property that either the
adjacency matrix A or the adjacency matrix A of the complement has smallest
eigenvalue at least −2. In this section we present other examples.
Recall that a distance-regular graph with diameter d has d+1 distinct eigen-
values and that its (adjacency) spectrum can be obtained from the intersection
array. Conversely, the spectrum of a distance-regular graph determines the in-
tersection array (see e.g. [117]). However, in general the spectrum of a graph
doesn’t tell you whether it is distance-regular or not.
For d ≥ 3 we have constructed graphs cospectral with, but non-isomorphic
to H(d, d) in §13.2.2. Many more examples are given in [183] and [120].
In the theory of distance-regular graphs an important question is: ‘Which
graphs are determined by their intersection array?’ For many distance-regular
graphs this is known to be the case. Here we investigate in the cases where the
graph is known to be determined by its intersection array, whether in fact it is
already determined by its spectrum.
13.5.1 Strongly regular DS graphs

The spectrum of a graph Γ determines whether Γ is strongly regular. Indeed,
by Proposition 3.3.1 we can see whether Γ is regular. And a regular graph with
spectrum θ1 ≥ ... ≥ θn is strongly regular if and only if |{θi | 2 ≤ i ≤ n}| = 2.
(That is, a regular graph is strongly regular if and only if either it is con-
nected, and then has precisely three distinct eigenvalues: its valency and two
others, or it is the disjoint union aKℓ of a (a ≥ 2) complete graphs of size ℓ.)
Indeed, if Γ has valency k and all eigenvalues θi with i > 1 are in {r, s}, then
(A − rI)(A − sI) = cJ so that A2 is a linear combination of A, I and J, and Γ
is strongly regular.
By Propositions 13.4.4 and 13.4.5 and Theorem 13.4.11, we find the following
infinite families of strongly regular DS graphs.
Proposition 13.5.1 If n 6= 8 and m 6= 4, the graphs aKℓ , L(Kn ) and L(Km,m )

and their complements are strongly regular DS graphs.
Note that L(Kn ) is the triangular graph T (n), and L(Km,m ) is the lattice
graph L2 (n). For n = 8 and m = 4 cospectral graphs exist. There is exactly
one graph cospectral with L(K4,4 ), the Shrikhande graph ([283]), and there are
three graphs cospectral with L(K8 ), the so-called Chang graphs ([80]). See also
§8.2.
Besides the graphs of Proposition 13.5.1, only a few strongly regular DS
graphs are known; these are surveyed in Table 13.2. (Here a local graph of a
graph Γ is the subgraph induced by the neighbors of a vertex of Γ.)
v spectrum name reference

√
5 2 [(−1 ± √ 5)/2]2 pentagon
13 6 [(−1 ± √13)/2]6 Paley [275]
17 8 [(−1 ± 17)/2]8 Paley [275]
16 5 110 (−3)5 folded 5-cube [274]
20
27 10 1 (−5)6 GQ(2,4) [274]
50 7 228 (−3)21 Hoffman-Singleton [183]
56 10 235 (−4)20 Gewirtz [152], [52]
77 16 255 (−6)21 M22 [42]
81 20 260 (−7)20 Brouwer-Haemers [51], §8.7
100 22 277 (−8)22 Higman-Sims [152]
105 32 284 (−10)20 flags of PG(2,4) [125]
112 30 290 (−10)21 GQ(3, 9) [76]
120 42 299 (−12)20 001... in S(5, 8, 24) [125]
126 50 2105 (−13)20 Goethals [97]
162 56 2140 (−16)21 local McLaughlin [76]
176 70 2154 (−18)21 01... in S(5, 8, 24) [125]
275 112 2252 (−28)22 McLaughlin [161]
Table 13.2: The known sporadic strongly regular DS graphs (up to comple-
ments)
Being DS seems to be a very strong property for strongly regular graphs.

Most strongly regular graphs have (many) cospectral mates. For example, there
are exactly 32548 non-isomorphic strongly regular graphs with spectrum 15, 315 ,
(−3)20 (cf. [244]). Other examples can be found in the survey [45]. One might
be tempted to conjecture that there are only finitely many strongly regular DS
graphs besides the ones from Proposition 13.5.1.
13.5.2 Distance-regularity from the spectrum

If d ≥ 3 only in some special cases does it follow from the spectrum of a graph
that it is distance-regular. The following result surveys the cases known to us.
Theorem 13.5.2 If Γ is a distance-regular graph with diameter d and girth g

satisfying one of the following properties, then every graph cospectral with Γ is
also distance-regular, with the same parameters as Γ.
(i) g ≥ 2d − 1,
(ii) g ≥ 2d − 2 and Γ is bipartite,
(iii) g ≥ 2d − 2 and cd−1 cd < −(cd−1 + 1)(θ1 + . . . + θd ),

(iv) Γ is a generalized Odd graph, that is, a1 = . . . = ad−1 = 0, ad 6= 0,
(v) c1 = . . . = cd−1 = 1,
(vi) Γ is the dodecahedron, or the icosahedron,
(vii) Γ is the coset graph of the extended ternary Golay code,
(viii) Γ is the Ivanov-Ivanov-Faradjev graph.
For parts (i), (iv) and (vi), see [52] (and also [178]), [201], and [183], respectively.
Parts (ii), (iii), (v), (vii) are proved in [117] (in fact, (ii) is a special case of (iii))
and (viii) is proved in [120]. Notice that the polygons Cn and the strongly
regular graphs are special cases of (i), while bipartite distance-regular graphs
with d = 3 (these are the incidence graphs of symmetric block designs, see also
[106, Thm.6.9]) are a special case of (ii).
An important result on spectral characterisations of distance-regular graphs
is the following theorem of Fiol & Garriga [143], a direct consequence of Theorem
11.10.1.
Theorem 13.5.3 Let Γ be a distance-regular graph with diameter d and kd =

|Γd (u)| vertices at distance d from any given vertex u. If Γ′ is cospectral with Γ
and |Γ′d (x)| = kd for every vertex x of Γ′ , then Γ′ is distance-regular.
Let us illustrate the use of this theorem by proving case (i) of Theorem 13.5.2.
Since the girth and the degree follow from the spectrum, any graph Γ′ cospectral
with Γ also has girth g and degree k1 . Fix a vertex x in Γ′ . Clearly cx,y = 1 for
every vertex y at distance at most (g − 1)/2 from x, and ax,y = 0 (where ax,y is
the number of neighbors of y at distance d(x, y) from x) if the distance between
x and y is at most (g − 2)/2. This implies that the number ki′ of vertices at
distance i from x equals k1 (k1 − 1)i−1 for i = 1, . . . , d − 1. Hence ki′ = ki for
these i. But then also kd′ = kd and Γ′ is distance-regular by Theorem 13.5.3.
13.5.3 Distance-regular DS graphs

Brouwer, Cohen & Neumaier [48] gives many distance-regular graphs deter-
mined by their intersection array. We only need to check which ones satisfy one
of the properties of Theorem 13.5.2. First we give the known infinite families:
Proposition 13.5.4 The following distance-regular graphs are DS.

(i) The polygons Cn .
(ii) The complete bipartite graphs minus a perfect matching.
(iii) The Odd graphs Od+1 .
(iv) The folded (2d + 1)-cubes.
As mentioned earlier, part (i) follows from property (i) of Theorem 13.5.2 (and
from Proposition 13.4.4). Part (ii) follows from property (ii) of Theorem 13.5.2,
and the graphs of parts (iii) and (iv) are all generalized Odd graphs, so the
result follows from property (iv), due to Huang & Liu [201].
Next, there are the infinite families where the spectrum determines the com-
binatorial or geometric structure, where the graphs are DS if and only if the
corresponding structure is determined by its parameters.
Proposition 13.5.5 (i) A graph cospectral with the incidence graph of a sym-
metric block design with parameters 2-(v, k, λ) is itself the incidence graph of a
symmetric block design with these same parameters.
In case (i) the designs known to be uniquely determined by their parameters

are the six projective planes P G(2, q) for q = 2, 3, 4, 5, 7, 8, and the biplane 2-
(11,5,2), and their complementary designs with parameters 2-(v, v−k, v−2k+λ).
The remaining known distance-regular DS graphs are presented in Tables 13.3,
13.4, 13.5. For all but one graph the fact that they are unique (that is, de-
termined by their parameters) can be found in [48]. Uniqueness of the Perkel
graph has been proved only recently [96]. The last columns in the tables refer to
the relevant theorems by which distance-regularity follows from the spectrum.
In these tables we denote by IG(v, k, λ) the point-block incidence graph of a
2-(v, k, λ) design, and by GH, GO, and GD the point graph of a generalized
hexagon, generalized octagon, and generalized dodecagon, respectively.
Recall that the point graph of a GH(1, q) (GO(1, q), GD(1, q)) is the point-
line incidence graph of a projective plane (generalized quadrangle, generalized
hexagon) of order q. Recall that the point graph of a GH(q, 1) is the line graph
of the dual GH(1, q), that is, the line graph of the point-line incidence graph
(also known as the flag graph) of a projective plane of order q.
Finally, G23 , G21 and G12 denote the binary Golay code, the doubly truncated bi-
nary Golay code and the extended ternary Golay code, and HoSi is the Hoffman-
Singleton graph.
n spectrum g name Thm.

√ 3 √
12 5 5 (−1)5 (− 5)3 3 icosahedron 13.5.2vi
√ 6 √
14 3 2 (− 2)6 (−3)1 6 Heawood; GH(1, 2) 13.5.2i
√ 6 √
14 4 2 (− 2)6 (−4)1 4 IG(7, 4, 2) 13.5.2ii
15 4 25 (−1)4 (−2)5 3 L(Petersen) 13.4.11
√ 6 √
21 4 (1+ 2) (1− 2)6 (−2)8 3 GH(2, 1) 13.5.2v
√ 10 √ 10
22 5 3 (− 3) (−5)1 4 IG(11, 5, 2) 13.5.2ii
√ 10 √
22 6 3 (− 3)10 (−6)1 4 IG(11, 6, 3) 13.5.2ii
√ 12 √
26 4 3 (− 3)12 (−4)1 6 GH(1, 3) 13.5.2i
√ 12 √
26 9 3 (− 3)12 (−9)1 4 IG(13, 9, 6) 13.5.2ii
36 5 216 (−1)10 (−3)9 5 Sylvester 13.5.2i
21
42 6 2 (−1)6 (−3)14 5 antipodal 6-cover of K7 13.5.2i
20
42 5 2 (−2)20 (−5)1 6 GH(1, 4) 13.5.2i
20
42 16 2 (−2)20 (−16)1 4 IG(21, 16, 12) 13.5.2ii
√ 12 √
52 6 (2+ √3) (2− √3)12 (−2)27 3 GH(3, 1) 13.5.2v
57 6 ( 3+2 5 )18 ( 3−2 5 )18 (−3)20 5 Perkel 13.5.2i
√ 30 √
62 6 5 (− 5)30 (−6)1 6 GH(1, 5) 13.5.2i
√ 30 √
62 25 5 (− 5)30 (−25)1 4 IG(31, 25, 20) 13.5.2ii
√ 27 √
63 8 8 (−1)8 (− 8)27 4 antipodal 7-cover of K9 13.5.2v
20
105 8 5 120 (−2)64 3 GH(4, 1) 13.5.2v
√ 56 √ 56
114 8 7 (− 7) (−8)1 6 GH(1, 7) 13.5.2i
√ 56 √
114 49 7 (− 7)56 (−49)1 4 IG(57, 49, 42) 13.5.2ii
continued...
n spectrum g name Thm.

√ 72 √
146 9 8 (− 8)72 (−9)1 6 GH(1, 8) 13.5.2i
√ 72 √
146 64 8 (− 8)72 (−64)1 4 IG(73, 64, 56) 13.5.2ii
175 21 728 221 (−2)125 3 L(HoSi) 13.4.11
√ 30 √ 30
186 10 (4+ 5) (4− 5) (−2)125 3 GH(5, 1) 13.5.2v
√ √
456 14 (6+ 7)56 (6− 7)56 (−2)343 3 GH(7, 1) 13.5.2v
506 15 4230 (−3)253 (−8)22 5 M23 graph 13.5.2i
512 21 5210 (−3)280 (−11)21 4 Coset graph of G21 13.5.2iii
√ 72 √
657 16 (7+ 8) (7− 8)72 (−2)512 3 GH(8, 1) 13.5.2v
729 24 6264 (−3)440 (−12)24 3 Coset graph of G12 13.5.2vii
819 18 5324 (−3)468 (−9)26 3 GH(2, 8) 13.5.2v
2048 23 7506 (−1)1288 (−9)253 4 Coset graph of G23 13.5.2iii, iv
2457 24 11324 3468 (−3)1664 3 GH(8, 2) 13.5.2v
Table 13.3: Sporadic distance-regular DS graphs with diameter 3

By Biaff(q) we denote the point-line incidence graph of an affine plane of order
q minus a parallel class of lines (sometimes called a biaffine plane). Any graph
cospectral with a graph Biaff(q) is also such a graph. For prime powers q < 9
there is a unique affine plane of order q. (Biaff(2) is the 8-gon.)
n nonnegative spectrum d g name Thm.
√ 6
18 31 3 04 4 6 Pappus; Biaff(3) 13.5.2ii
30 31 29 010 4 8 Tutte’s 8-cage; GO(1, 2) 13.5.2i
32 41 212 06 4 6 Biaff(4) 13.5.2ii
√ 20
50 51 5 08 4 6 Biaff(5) 13.5.2ii
√ 24
80 41 6 030 4 8 GO(1, 3) 13.5.2i
√ 42
98 71 7 012 4 6 Biaff(7) 13.5.2ii
1
√ 21 √ 27 28
126 3 6 2 0 6 12 GD(1, 2) 13.5.2i
1
√ 56 14
128 8 8 0 4 6 Biaff(8) 13.5.2ii
√ 50
170 51 8 068 4 8 GO(1, 4) 13.5.2i
Table 13.4: Sporadic bipartite distance-regular DS graphs with d ≥ 4

n spectrum d g name Thm.
1
√ 3 √
20 3 5 1 0 (−2) (− 5)3
5 4 4
5 5 dodecahedron 13.5.2vi
√ √
28 31 28 (−1+ 2)6 (−1)7 (−1− 2)6 4 7 Coxeter 13.5.2i
45 41 39 1√10 (−1)9 (−2)16 4 3 GO(2, 1) 13.5.2v
102 31 ( 1+2 17 )9√ 218 θ116 017 7 9 Biggs-Smith 13.5.2v
θ216 ( 1−2 17 )9 θ316
(θ1 , θ2 , θ3 roots of θ3 + 3θ2 − 3 = 0)
√ √
160 61 (2+ 6)24 230 (2− 6)24 (−2)81 4 3 GO(3, 1) 13.5.2v
√ √
189 41 (1+ 6)21 (1+ 2)27 128 6 3 GD(2, 1) 13.5.2v
√ √
(1− 2)27 (1− 6)21 (−2)64
330 71 455 1154 (−3)99 (−4)21 4 5 M22 graph 13.5.2v
√ √
425 8 (3 + 8)50 368 √
1
(3 − 8)50 (−2)256 4 3 GO(4, 1) 13.5.2v
990 71 542 455 ( −1+2√ 33 )154 1154 0198 8 5 Ivanov-Ivanov-
(−3)99 ( −1−2 33 )154 (−4)21 Faradjev 13.5.2viii
Table 13.5: Sporadic non-bipartite distance-regular DS graphs with d ≥ 4

We finally remark that also the complements of distance-regular DS graphs
are DS (but not distance-regular, unless d = 2).
13.6 The method of Wang & Xu

Wang & Xu [301] invented a method to show that relatively many graphs are
determined by their spectrum and the spectrum of their complement. A sketch.
Let Γ be a graph on n vertices with adjacency matrix A. The walk matrix W
of Γ is the square matrix of order n with i-th column Ai−1 1 (1 ≤ i ≤ n). It is
nonsingular if and only if A does not have an eigenvector orthogonal to 1.
(Indeed, let u⊤ A = θu⊤ . Then u⊤ W = (1, θ, ..., θn−1 )u⊤ 1. If u⊤ 1 = 0 then
this shows that the rows of W are dependent. If for no eigenvector u⊤ we have
u⊤ 1 = 0, then all eigenvalues have multiplicity 1, and by Vandermonde W is
nonsingular.)
P i
Let p(t) = ci t = det(tI − A) be the characteristic polynomial of A. Let
the companion matrix C = (cij ) be given by cin = −ci and cij = δi,j+1 for
1 ≤ j ≤ n − 1. Then AW = W C.
(Indeed, this follows from p(A) = 0.)
Assume that Γ and Γ′ are cospectral with cospectral complements. Call their
walk matrices W and W ′ . Then W ⊤ W = W ′⊤ W ′ .
(Indeed, (W ⊤ W )i,j = 1⊤ Ai+j−2 1, and we saw in the proof of Proposition
13.1.1 that if Γ and Γ′ , with adjacency matrices A and A′ , are y-cospectral for
two distinct y, then 1⊤ Am 1 = 1⊤ A′m 1 for all m.)
Suppose that W is nonsingular. Then W ′ is nonsingular, and Q = W ′ W −1 is
the unique orthogonal matrix such that A′ = QAQ⊤ and Q1 = 1.
(Indeed, since W ⊤ W = W ′⊤ W ′ also W ′ is nonsingular, and Q1 = 1
since QW = W ′ , and QQ⊤ = W ′ (W ⊤ W )−1 W ′⊤ = I. Since Γ and Γ′ are
cospectral, their companion matrices are equal and QAQ⊤ = QW CW −1 Q⊤ =
W ′ CW ′−1 = A′ . If Q is arbitrary with QQ⊤ = I, Q1 = 1 (hence also Q⊤ 1 = 1)
and QAQ⊤ = A′ , then QAm 1 = QAm Q⊤ 1 = A′m 1 for all m, and QW = W ′ .)
Forget about Γ′ and study rational matrices Q with QQ⊤ = I, Q1 = 1 and
QAQ⊤ a (0,1)-matrix with zero diagonal. Let the level of Q be the smallest
integer ℓ such that ℓQ is integral. The matrices Q of level 1 are permutation
matrices leading to isomorphic graphs. So the graph Γ (without eigenvector
orthogonal to 1) is determined by its spectrum and the spectrum of its comple-
ment when all such matrices Q have level 1.
If Q has level ℓ, then clearly ℓ| det W . A tighter restriction on ℓ is found by
looking at the Smith Normal Form S of W . Let S = U W V with unimodular
integral U and V , where S = diag(s1 , ..., sn ) with s1 |s2 |...|sn . Then W −1 =
V S −1 U so that sn W −1 is integral, and ℓ|sn .
Let p be prime, p|ℓ. There is an integral row vector z, z 6≡ 0 (mod p) such that
zW ≡ 0 (mod p) and zz ⊤ ≡ 0 (mod p).
(Indeed, let z be a row of ℓQ, nonzero mod p. Now QW = W ′ is integral
and hence zW ≡ 0 (mod p). And QQ⊤ = I, so zz ⊤ = ℓ2 ≡ 0 (mod p).)
This observation can be used to rule out odd prime divisors of ℓ in some cases.
Suppose that all numbers si are powers of 2, except possibly the last one sn .
Let p be an odd prime divisor of sn , and suppose that uu⊤ 6≡ 0 (mod p), where
u is the last row of U . Then p ∤ ℓ.
(Indeed, zW ≡ 0 (mod p) and W = U −1 SV −1 with unimodular V implies
zU S ≡ 0 (mod p). Assume p|ℓ, so that p|sn . Let y = zU −1 . Then all
−1
13.7. EXERCISES 193
coordinates of y except for the last one are 0 (mod p). And z = yU is a nonzero
constant times u (mod p). This contradicts uu⊤ 6≡ 0 (mod p).)
It remains to worry about p = 2. Assume that sn ≡ 2 (mod 4), so that (with
all of the above assumptions) ℓ = 2. For z we now have z 6≡ 0 (mod 2), zW ≡ 0
(mod 2), zz ⊤ = 4, z1 = 2, so that z has precisely four nonzero entries, three 1
and one −1.
We proved the following:
Theorem 13.6.1 Let Γ be a graph on n vertices without eigenvector orthogonal

to 1, and let S = diag(s1 , ..., sn ) = U W V be the Smith Normal Form of its walk
matrix W , where U and V are unimodular. Let u be the last row of U . If sn = 2
(mod 4), and gcd(uu⊤ , sn /2) = 1, and zW 6= 0 (mod 2) for every (0, 1)-vector
z with weight 4, then Γ is determined by its spectrum and the spectrum of its
complement.
Wang & Xu generate a number of random graphs where this method applies.
Let us abbreviate the condition ‘determined by its spectrum and the spec-
trum of its complement’ by DGS (determined by the generalized spectrum).
Wang & Xu [302] used their approach to find conditions for which a DGS graph
remains DGS if an isolated vertex is added.
Theorem 13.6.2 Let Γ be a graph without eigenvector orthogonal to 1. If

we have gcd(det A, det W ) = 1, then the graph obtained from Γ by adding an
isolated vertex is DGS if and only if Γ is.
There is experimental evidence that in most cases where a cospectral mate

exists, the level ℓ is 2.
13.7 Exercises
Exercise 1 Show for the adjacency matrix A
(i) that there is no pair of cospectral graphs on fewer than 5 vertices,
(ii) that the Saltire pair is the only cospectral pair on 5 vertices,
(iii) that there are precisely 5 cospectral pairs on 6 vertices.
s s
HsH
s s s s s s s s s s s s s
s s s s s s s s s s s s s s
s s s
sH
s s s s s s s sHs s sHss
s Hs
@@
s s s s s s s
s Hs sHs
x6 − x4 x6 − 4x4 + 3x2 x6 − 5x4 + 4x2 x6 − 6x4 − 4x3 x6 − 7x4 − 4x3

+ 5x2 + 4x + 7x2 + 4x − 1
Table 13.6: The cospectral graphs on 6 vertices (with char. pol.)

Chapter 14
Graphs with few

eigenvalues
Graphs with few distinct eigenvalues tend to have some kind of regularity. A
graph with only one eigenvalue (for A or L or Q) is edgeless, and a connected
graph with two distinct adjacency eigenvalues (for A or L or Q) is complete. A
connected regular graph Γ has three eigenvalues if and only if Γ is connected
and strongly regular. Two obvious next cases are connected regular graphs with
four eigenvalues, and general graphs with three eigenvalues. In the latter case
the graphs need not be regular, so it matters which type of matrix we consider.
For the Laplacian matrix there is an elegant characterization in terms of the
structure, which gives a natural generalization of the spectral characterization
of strongly regular graphs.
14.1 Regular graphs with four eigenvalues

Suppose Γ is regular with r distinct (adjacency) eigenvalues k = λ1 > . . . > λr .
Then the Laplacian matrix has eigenvalues 0 = k − λ1 < . . . < k − λr , and
the signless Laplacian has eigenvalues k + λ1 > . . . > k + λr . So for regular
graphs these three matrices have the same number of distinct eigenvalues. If, in
addition, both Γ and its complement Γ are connected, then Γ also has r distinct
eigenvalues, being n − k − 1 > −λr − 1 > . . . > −λ2 − 1. However, for the Seidel
matrix the eigenvalues become −2λr − 1 > . . . > −2λ2 − 1 and n − 2k − 1. But
n − 2k − 1 may be equal to one of the other eigenvalues in which case S has r − 1
distinct eigenvalues. For example, the Petersen has three distinct adjacency
eigenvalues, but only two distinct Seidel eigenvalues, being ±3.
Connected regular graphs with four distinct (adjacency) eigenvalues have
been studied by Doob [132, 133], Van Dam [112], and Van Dam & Spence [122].
Many such graphs are known, for example the line graphs of primitive strongly
regular graphs, and distance regular graphs of diameter 3. More generally,
most graphs defined by a relation of a three-class association scheme have four
eigenvalues. There is no nice characterization as for regular graphs with three
eigenvalues, but they do possess an interesting regularity property. A graph is
walk-regular, whenever for every ℓ ≥ 2 the number of closed walks of length ℓ at
a vertex v is independent of the choice of v. Note that walk-regularity implies
195
196 CHAPTER 14. GRAPHS WITH FEW EIGENVALUES
regularity (take ℓ = 2). Examples of walk-regular graphs are distance-regular

graphs, and vertex-transitive graphs, but there is more.
Proposition 14.1.1 Let Γ be a connected graph whose adjacency matrix A has

r ≥ 4 distinct eigenvalues. Then Γ is walk-regular if and only if Aℓ has constant
diagonal for 2 ≤ ℓ ≤ r − 2.
Proof. We know that the number of closed walks of length ℓ at vertex v

equals (Aℓ )v,v . Therefore, Γ is walk-regular if and only if Aℓ has constant
diagonal for all ℓ ≥ 2. Suppose Aℓ has constant diagonal for 2 ≤ ℓ ≤ r − 2.
Then A2 has constant diagonal, so Γ is
regular. The Hoffman polynomial of
r−1 r−2 2
Γ has
degree r − 1, and hence A ∈ A , . . . , A , A, I, J . This implies
Aℓ ∈ Ar−2 , . . . , A2 , A, I, J for all ℓ ≥ 0. Therefore Aℓ has constant diagonal
for all ℓ ≥ 0.
Corollary 14.1.2 If Γ is connected and regular with four distinct eigenvalues,
then Γ is walk-regular.
For a graph Γ with adjacency matrix A, the average number of triangles through
1
a vertex equals 2n trA3 . Suppose Γ is walk-regular. Then this number must be
1
an integer. Similarly, 2n trAℓ is an integer if ℓ is odd, and n1 trAℓ is an integer
if n is even. Van Dam and Spence [122] have used these (and other) conditions
in their computer generation of feasible spectra for connected regular graphs
with four eigenvalues. For constructions, characterizations, and other results
on regular graphs with four eigenvalues we refer to Van Dam [112, 113]. Here
we finish with the bipartite case, which can be characterized in terms of block
designs (see §4.7).
Proposition 14.1.3 A connected bipartite regular graph Γ with four eigen-
values is the incidence graph of a symmetric 2-design (and therefore distance-
regular).
Proof. Since Γ is connected, bipartite and regular the spectrum is
{k, λ2v−1 , (−λ2 )v−1 , −k} ,
where 2v is the number of vertices. For the adjacency matrix A of Γ, we have

O N 2 NN⊤ O
A= , and A = ,
N⊤ O O N ⊤N
for some square (0, 1)-matrix N satisfying N 1 = N ⊤ 1 = k1. It follows that

N N ⊤ has spectrum {(k 2 )1 , (λ22 )v−1 }, where k 2 corresponds to the row and col-
umn sum of N N ⊤ . This implies that N N ⊤ ∈ hJ, Ii, and hence N is the
incidence matrix of a symmetric design.
14.2 Three Laplacian eigenvalues

If a connected graph Γ has three distinct Laplacian eigenvalues 0 < ν < ν ′ (say),
the complement Γ has eigenvalues 0 ≤ n − ν ′ < n − ν, so if Γ is connected, it
also has three distinct eigenvalues. To avoid the disconnected exceptions, it is
convenient to use the notion of restricted eigenvalues (recall that an eigenvalue
14.2. THREE LAPLACIAN EIGENVALUES 197
is restricted if it has an eigenvector orthogonal to the all-one vector 1), and

consider graphs with two distinct restricted Laplacian eigenvalues.
We say that a graph Γ has constant µ(Γ) if Γ is not complete and any two
distinct nonadjacent vertices of Γ have the same number of common neighbors
(equal to µ(Γ)).
Theorem 14.2.1 A graph Γ has two distinct restricted Laplacian eigenvalues
ν and ν ′ if and only if Γ has constant µ(Γ) and its complement Γ has constant
µ(Γ). If Γ is such a graph, only two vertex degrees d and d′ occur, and
ν + ν ′ = d + d′ + 1 = µ(Γ) + n − µ(Γ), νν ′ = dd′ + µ(Γ) = µ(Γ)n .
Proof. Suppose Γ has just two restricted Laplacian eigenvalues ν and ν ′ . Then
(L − νI)(L − ν ′ I) has rank 1 and row sum νν ′ , so
νν ′
(L − νI)(L − ν ′ I) = J.
n
If u and v are nonadjacent vertices, then (L)uv = 0, so (L2 )uv = νν ′ /n, and
µ(Γ) = νν ′ /n is constant. Similarly, Γ has constant µ(Γ) = (n − ν)(n − ν ′ )/n.
Next suppose µ = µ(Γ) and µ = µ(Γ) are constant. If u and v are adjacent
vertices, then ((nI − J − L)2 )uv = µ, so µ = (L2 )uv + n, and if u and v are
nonadjacent, then (L2 )uv = µ. Furthermore (L2 )uu = d2u + du , where du is the
degree of u. Writing D = diag(d1 , . . . , dn ), we obtain
L2 = (µ − n)(D − L) + µ(J − I − D + L) + D2 + D =
(µ + n − µ)L + D2 − (µ + n − µ − 1)D − µI + µJ .
Since L and L2 have zero row sums, it follows that du 2 − du (µ + n − µ − 1) − µ +

µn = 0 for every vertex u. So L2 − (µ + n − µ)L + µnI = µJ. Now let ν and ν ′
′
be such that ν + ν ′ = µ + n − µ and νν ′ = µn, then (L − νI)(L − ν ′ I) = νν n J,
so L has distinct restricted eigenvalues ν and ν ′ . As a side result we obtained
that all vertex degrees du satisfy the same quadratic equation, so that du can
only take two values d and d′ , and the formulas readily follow.
Regular graphs with constant µ(Γ) and µ(Γ) are strongly regular. So Theo-
rem 14.2.1 generalizes the spectral characterization of strongly regular graphs.
Several nonregular graphs with two restricted Laplacian eigenvalues are known.
A geodetic graph of diameter three with connected complement provides an
example with µ(Γ) = 1 (see [48], Theorem 1.17.1). Here we give two other
constructions. Both constructions use symmetric block designs (see §4.7). Cor-
rectness easily follows by use of Theorem 14.2.1.
Proposition 14.2.2 Let N be the incidence matrix of a symmetric 2-(n, k, λ)
design. Suppose that N is symmetric (which means that the design has a polar-
ity). Then L = kI − N√ is the Laplacian matrix of a graph with two restricted
eigenvalues, being k ± k − λ. The possible degrees are k and k − 1.
If all diagonal elements of N are 0, then the graph Γ is a (n, k, λ)-graph (a
strongly regular graph with λ = µ), and if all diagonal elements of N are 1,
then Γ is such a graph. Otherwise both degrees k and k − 1 do occur. For
example the Fano plane admits a symmetric matrix with three ones on √ the
diagonal. The corresponding graph has restricted Laplacian eigenvalues 3 ± 2,
and vertex degrees 2 and 3. See also §4.8.
Proposition 14.2.3 Let N be the incidence matrix of a symmetric block de-

sign. Write
 
vI − J O N1 − J
1 N1
N= , and define L =  O vI − J −N2 .
0 N2 ⊤ ⊤
N1 − J −N2 2(k − λ)I
Then L has two restricted eigenvalues.
Other examples, characterizations and a table of feasible spectra can be found

in [116] and [113] (see also Exercise 1). See [303] for some more recent results
on graphs with three Laplacian eigenvalues.
14.3 Other matrices with at most three eigen-

values
No characterization is known of nonregular graphs with three M -eigenvalues, for
a matrix M other than the Laplacian. However several examples and properties
are known. Some of these will be discussed below.
14.3.1 Few Seidel eigenvalues

Seidel switching (see §1.8.2) doesn’t change the Seidel spectrum, so having few
Seidel eigenvalues is actually a property of the switching class of a graph. For
example the switching class of Kn , the edgeless graph on n vertices, consists of
the complete bipartite graphs Km,n−m , and all of them have Seidel spectrum
{(−1)n−1 , n − 1}. Only the one-vertex graph K1 has one Seidel eigenvalue.
Graphs with two Seidel eigenvalues are strong (see §9.1). To be precise, they
are the graphs whose associated two-graph is regular (Theorem 9.3.1). The
Seidel matrix is a special case of a generalized adjacency matrix. These are
matrices of the form M (x, y, z) = xI + yA + z(J − I − A) with y 6= z, where A
is the adjacency matrix; see also Chapter 13. If A is the adjacency matrix of a
strongly regular graph with eigenvalues k ≥ r > s, then both nA − (k − r)J and
nA − (k − s)J (these are basically the nontrivial idempotents of the association
scheme) are generalized adjacency matrices with two eigenvalues. We recall that
a strong graph either has two Seidel eigenvalues, or is strongly regular. Thus
for every strong graph there exist numbers x, y and z, such that M (x, y, z) has
two eigenvalues.
Proposition 14.3.1 A graph is strong if and only if at least one generalized

adjacency matrix has two eigenvalues.
Proof. Correctness of the ‘only if’ part of the statement has been established
already. Without loss of generality we assume that the eigenvalues of M =
M (x, y, z) are 0 and 1. So M satisfies M 2 = M . Let di be the degree of
vertex i, then x = Mii = (M 2 )ii = x2 + di y 2 + (n − 1 − di )z 2 , which gives
di (y 2 − z 2 ) = x − x2 − (n − 1)z 2 . So y = −z or Γ is regular. In the first case
S = z1 (M − xI) is the Seidel matrix of Γ with two eigenvalues, so Γ is strong. In
1
case Γ is regular, the adjacency matrix A = y−z (M + (z − x)I − zJ) has three
eigenvalues, so Γ is strongly regular and therefore strong.
14.3. OTHER MATRICES WITH AT MOST THREE EIGENVALUES 199
So if a generalized adjacency matrix M (x, y, z) of a nonregular graph has

two eigenvalues, then y = −z (and we basically deal with the Seidel matrix).
A strongly regular graph Γ on n vertices with adjacency eigenvalues k, r,
s (k ≥ r > s) has Seidel eigenvalues ρ0 = n − 1 − 2k, ρ1 = −2s − 1, and
ρ2 = −2r − 1. If ρ0 = ρ1 , or ρ0 = ρ2 , then Γ has two eigenvalues, otherwise Γ,
and all graphs switching equivalent to Γ, have three eigenvalues. For example,
the (switching class of the) Petersen graph has two Seidel eigenvalues
√ 3 and −3,
whilst the pentagon C5 has three Seidel eigenvalues 0 and ± 5. However, not
every graph with three Seidel eigenvalues is switching equivalent to a strongly
regular graph. Not even if the graph is regular. Indeed, consider a graph Γ
whose Seidel matrix S has two eigenvalues ρ1 and ρ2 . Then (S + I) ⊗ (S + I) − I
represents a graph Γ2 with eigenvalues (ρ1 + 1)2 − 1, (ρ1 + 1)(ρ2 + 1) − 1, and
(ρ2 + 1)2 − 1. Moreover, Γ2 is regular if Γ is.
14.3.2 Three adjacency eigenvalues

Connected regular graphs with three adjacency eigenvalues√are strongly
√ regular.
The complete bipartite graphs Kℓ,m have spectrum {− ℓm, 0n−2 , ℓm}. If
ℓ 6= m they are nonregular with three adjacency eigenvalues. Other nonregular
graphs with three adjacency eigenvalues have been constructed by Bridges and
Mena [39], Klin and Muzychuk [217], and Van Dam [113, 114]. Chuang and
Omidi [84] characterized all such graphs with largest eigenvalue at most eight.
Many nonregular graphs with three eigenvalues can be made from a strongly
regular graphs by introducing one new vertex adjacent to all other vertices.
Such a graph is called a cone over a strongly regular graph.
Proposition 14.3.2 Let Γ be a strongly regular graph on n vertices with eigen-

values k > r > s. Then the cone Γ̂ over Γ has three eigenvalues if and only if
n = s(s − k).
Proof. If Â is the adjacency matrix of Γ̂, then Â admits an equitable partition

with quotient matrix
0 n
1 k
√
with eigenvalues (k ± k 2 + 4n)/2, which are also eigenvalues of Â. The other
eigenvalues of Â have eigenvectors orthogonal to the characteristic vectors of
the partition, so they remain eigenvalues if the all-one blocks of the equitable
partition are replaced by all-zero blocks. Therefore they are precisely
√ the re-
stricted eigenvalues r and s of Γ. So the eigenvalues of Â are (k ± k 2 + 4n)/2,
√
r and s. Two of these values coincide if and only if s = (k − k 2 + 4n)/2.
There exist infinitely many strongly regular graphs for which n = s(s − k),
the smallest of which is the Petersen graph. The cone over the Petersen graph
has eigenvalues 5, 1 and −2. If a cone over a strongly regular graph has three
eigenvalues, then these eigenvalues are integers (see Exercise 3). The complete
bipartite graphs provide many examples with nonintegral eigenvalues. In fact:
Proposition 14.3.3 If Γ is a connected graph with three distinct adjacency

eigenvalues of which the largest is not an integer, then Γ is a complete bipartite
graph.
Proof. Assume Γ has n ≥ 4 vertices. Since the largest eigenvalue ρ is non-

integral with multiplicity 1, one of the other two eigenvalues ρ (say) also has
this property, and the third eigenvalue has multiplicity n − 2 ≥ 2, so cannot be
irrational. Thus the spectrum of Γ is
√ √
{ρ = 12 (a + b), ρ = 12 (a − b), cn−2 },
for integer a, b and c. Now trA = 0 gives c = −a/(n − 2). By Perron-Frobenius’

theorem,
√ ρ ≥ | ρ |, therefore a ≥ 0 and c ≤ 0. If c = 0, the eigenvalues of Γ are
± b/2 and 0, and Γ is bipartite of diameter at most 2, and hence Γ is complete
bipartite. If c ≤ −2, then trA2 ≥ 4(n − 2)2 so Γ has 2
√ at least 2(n − 2)√ edges
1
which is ridiculous. If c = −1, then ρ = 2 (n − 2 + b) ≤ n − 1, hence b ≤ n
and ρ > −1. This implies that A + I is positive semi-definite (of rank 2). So
A + I is the Gram matrix of a set of unit vectors (in R2 ) with angles 0 and
π/2. This implies that being adjacent is an equivalence relation, so Γ = Kn , a
contradiction.
The conference graphs are examples of regular graphs where only the largest
eigenvalue is an integer. Van Dam and Spence [72]√found a number of nonregular
graphs on 43 vertices with eigenvalues 21, − 12 ± 21 41. It turns out that all these
graphs have three distinct vertex degrees: 19, 26 and 35 (which was impossible
in case of the Laplacian spectrum).
14.3.3 Three signless Laplacian eigenvalues

Recently, Ayoobi, Omidi and Tayfeh-Rezaie [11] started to investigate nonreg-
ular graphs whose signless Laplacian matrix Q has three distinct eigenvalues.
They found three infinite families.
(i) The complete Kn with one edge deleted has Q-spectrum
1 p 1 p
{ (3n − 6 + n2 + 4n − 12), (n − 2)n−2 , (3n − 6 − n2 + 4n − 12)} .
2 2
(ii) The star K1,n−1 has Q-spectrum 01 , 1n−2 , n1 .

(iii) The complement of Km,m + mK1 has Q-spectrum
(5m − 2)1 , (3m − 2)m , (2m − 2)2m−2 .
In addition there are some sporadic examples (see also Exercise 4). Like in
Proposition 14.3.3 the case in which the spectral radius is nonintegral can be
characterized.
Proposition 14.3.4 [11] Let Γ be a connected graph on at least four vertices

of which the signless Laplacian has three distinct eigenvalues. Then the largest
of these eigenvalues is nonintegral if and only if Γ is the complete graph minus
one edge.
It is not known if there exist other nonregular examples with a nonintegral

eigenvalue. We expect that the above list is far from complete.
14.4. EXERCISES 201
14.4 Exercises
Exercise 1 Prove that a graph with two restricted Laplacian eigenvalues whose
degrees d and d′ differ by 1, comes from a symmetric design with a polarity as
described in Proposition 14.2.2.
Exercise 2 Let Γ be a strongly regular graph with a coclique C whose size
meets Hoffman’s bound (4.1.2). Prove that the subgraph of Γ induced by the
vertices outside C is regular with at most four distinct eigenvalues. Can it have
fewer than four eigenvalues?
Exercise 3 Suppose Γ̂ is a cone over a strongly regular graph. Show that, if Γ̂
has three distinct eigenvalues, then all three are integral.
Exercise 4 Show that the cone over the Petersen graph has three signless
Laplacian eigenvalues. Find a necessary and sufficient condition on the param-
eters (n, k, λ.µ) of a strongly regular graph Γ under which the cone over Γ has
three signless Laplacian eigenvalues.
Bibliography
[1] M. Ajtai, J. Komlós & E. Szemerédi, Sorting in c log n parallel steps,

Combinatorica 3 (1983) 1–19. (p. 68)
[2] N. Alon, Eigenvalues and Expanders, Combinatorica 6 (1986) 83–96.

(pp. 66, 67)
[3] N. Alon, The Shannon Capacity of a union, Combinatorica 18 (1998)

301–310. (p. 65)
[4] N. Alon, Large sets in finite fields are sumsets, J. Number Th. 126 (2007)
110–118. (p. 90)
[5] N. Alon & F. R. K. Chung, Explicit Constructions of linear sized tolerant

networks, Discr. Math. 2 (1988) 15–19. (p. 68)
[6] N. Alon & V. D. Milman, λ1 , isoperimetric inequalities for graphs and

superconcentrators, J. Combin. Th. (B) 38 (1985) 73–88. (p. 71)
[7] N. Alon & P. Seymour, A Counterexample to the rank-coloring conjecture,

J. Graph Theory 13 (1989) 523–525. (p. 64)
[8] W. N. Anderson & T. D. Morley, Eigenvalues of the Laplacian of a graph,

Lin. Multilin. Alg. 18 (1985) 141–145. (p. 48)
[9] D. M. Appleby, Symmetric informationally complete-positive operator val-

ued measures and the extended Clifford group, J. Math. Phys. 46 052107
(2005). (p. 143)
[10] M. Aschbacher, The non-existence of rank three permutation groups of

degree 3250 and subdegree 57, J. Algebra 19 (1971) 538–540. (p. 153)
[11] F. Ayoobi, G. R. Omidi & B. Tayfeh-Rezaie, A note on graphs whose

signless Laplacian has three distinct eigenvalues, preprint. (p. 200)
[12] Bhaskar Bagchi, On strongly regular graphs with µ ≤ 2, Discr. Math. 306
(2006) 1502–1504. (p. 106)
[13] Hua Bai, The Grone-Merris Conjecture, arXiv:0911.2172v3, 1 Jan 2010.

(p. 50)
[14] G. A. Baker, Drum shapes and isospectral graphs, J. Math. Phys. 7 (1966)
2238–2242. (p. 26)
203
204 BIBLIOGRAPHY
[15] S. Bang, A. Dubickas, J. H. Koolen & V. Moulton, There are only

finitely many distance-regular graphs of fixed valency greater than two,
arXiv.org:0909.5253, 29 Sep 2009. (p. 160)
[16] E. Bannai & T. Ito, On finite Moore graphs, J. Fac. Sci. Univ. Tokyo, Sect
IA 20 (1973) 191–208. (p. 161)
[17] E. Bannai & T. Ito, Algebraic Combinatorics I: Association Schemes,
Benjamin/Cummings, London, 1984. (p. 160)
[18] D. Bauer, H. J. Broersma & H. J. Veldman, Not every 2-tough graph is
Hamiltonian, Discr. Appl. Math. 99 (2000) 317–321. (p. 69)
[19] D. Bauer, H. J. Broersma, J. van den Heuvel & H. J. Veldman, Long
cycles in graphs with prescribed toughness and minimum degree, Discr.
Math. 141 (1995) 1–10. (p. 69)
[20] V. Belevitch, Theory of 2n-terminal networks with application to confer-
ence telephony, Electron. Comm. 27 (1950) 231–244. (p. 140)
[21] E. R. Berlekamp, J. H. van Lint & J. J. Seidel, A strongly regular graph
derived from the perfect ternary Golay code, pp. 25–30 in: A survey of
combinatorial theory, Symp. Colorado State Univ., 1971, (J.N. Srivastava
et al., eds.), North Holland, 1973. (p. 131)
[22] A. Berman & R. J. Plemmons, Nonnegative matrices in the mathematical
sciences, Acad. Press, New York, 1979. (p. 30)
[23] Jochem Berndsen & Mayank, The Brouwer conjecture for split graphs,
preprint, 2010. (p. 50)
[24] Monica Bianchini, Marco Gori & Franco Scarselli, Inside PageRank, ACM
Trans. on Internet Technology 5 (2005) 92–128. (p. 56)
[25] A. Bigalke & H. A. Jung, Über Hamiltonsche Kreise und unabhängige
Ecken in Graphen, Monatsh. Math. 88 (1979) 195–210. (p. 112)
[26] N. L. Biggs, Algebraic Graph Theory, Cambridge Tracts in Mathematics,
Vol 67, Cambridge Univ. Press, Cambridge, 1974. (Second ed., Cambridge,
1993.) (Not cited)
[27] A. Blokhuis, Few-distance sets, Ph.D. thesis, Eindhoven, 1983. (p. 145)
[28] A. Blokhuis & A. E. Brouwer, Uniqueness of a Zara graph on 126 points
and non-existence of a completely regular two-graph on 288 points, pp 6-19
in: Papers dedicated to J. J. Seidel, P. J. de Doelder, J. de Graaf & J. H.
van Lint (eds.), Eindhoven Univ. Techn. Report 84-WSK-03, Aug 1984.
Zbl 554.05049 (p. 139)
[29] J. A. Bondy, A graph reconstruction manual, pp. 221–252 in: Surveys in
Combinatorics, LMS Lecture Note Series 166, Cambridge Univ. Press,
Cambridge, 1991. (p. 24)
[30] J. Bosák, A. Rosa & Š. Znám, On decompostions of complete graphs into
factors with given diameters, pp. 37–56 in: Theory of Graphs (Proc. Col-
loq., Tihany, 1966), Academic Press, New York, 1968. (p. 20)
BIBLIOGRAPHY 205
[31] R. C. Bose, Strongly regular graphs, partial geometries and partially bal-
anced designs, Pacific J. Math 13 (1963) 389–419. (p. 109)
[32] R. C. Bose & T. A. Dowling, A generalization of Moore graphs of diameter

two, J. Combin. Th. 11 (1971) 213–226. (p. 131)
[33] R. C. Bose & D. M. Mesner, On linear associative algebras correspond-

ing to association schemes of partially balanced designs, Ann. Math.
Statist. 30 (1959) 21–38. (p. 148)
[34] R. C. Bose & T. Shimamoto, Classification and analysis of partially bal-

anced incomplete block designs with two associate classes, J. Am. Statist.
Assoc. 47 (1952) 151–184. (p. 147)
[35] N. Bourbaki, Groupes et algèbres de Lie, Chap. 4, 5 et 6, Hermann, Paris,

1968. (p. 41)
[36] V. Brankov, P. Hansen & D. Stevanović, Automated conjectures on upper

bounds for the largest Laplacian eigenvalue of graphs, Lin. Alg. Appl. 414
(2006) 407–424. (p. 48)
[37] A. Brauer & I. C. Gentry, On the characteristic roots of tournament ma-

trices, Bull. Amer. Math. Soc. 74 (1968) 1133–1135. (p. 25)
[38] W. G. Bridges & R. A. Mena, Rational circulants with rational spectra and
cyclic strongly regular graphs, Ars. Combin. 8 (1979) 143–161. (p. 104)
[39] W. G. Bridges & R. A. Mena, Multiplicative cones - a family of three

eigenvalue graphs, Aequationes Math. 22 (1981) 208–214. (p. 199)
[40] Sergey Brin & Lawrence Page, The anatomy of a large-scale hypertextual
web search engine, Proceedings of the Seventh International World Wide
Web Conference, Computer Networks and ISDN Systems 30 (1998) 107–
117. (p. 55)
[41] R. Brooks, Non-Sunada graphs, Ann. Inst. Fourier (Grenoble) 49 (1999)

707–725. (pp. 176, 179)
[42] A. E. Brouwer, The uniqueness of the strongly regular graph on 77 points,

J. Graph Theory 7 (1983) 455–461. (pp. 128, 188)
[43] A. E. Brouwer, Toughness and spectrum of a graph, Lin. Alg. Appl. 226–
228 (1995) 267–271. (p. 69)
[44] A. E. Brouwer, A non-degenerate generalized quadrangle with lines of size

four is finite, in: Advances in finite geometries and designs, Proc. 3rd Isle
of Thorns Conf., Chelwood Gate/UK 1990, pp. 47–49. (p. 115)
[45] A. E. Brouwer, Strongly regular graphs, in: The CRC handbook of combi-
natorial designs, (C.J. Colbourn and J.H. Denitz eds.), CRC Press, 1996,
pp. 667–685. (p. 188)
[46] A. E. Brouwer, Small integral trees, Electr. J. Combin. 15 (2008) N1.

(p. 85)
206 BIBLIOGRAPHY
[47] A. E. Brouwer & A. M. Cohen, Some remarks on Tits geometries, with

an appendix by J. Tits, Indag. Math. 45 (1983) 393–402. (p. 113)
[48] A. E. Brouwer, A. M. Cohen & A. Neumaier, Distance-Regular Graphs,

Springer-Verlag, 1989. (pp. 41, 104, 105, 106, 123, 157, 161, 176, 189, 190,
197)
[49] A. E. Brouwer & C. A. van Eijl, On the p-rank of the adjacency matrices
of strongly regular graphs, J. Alg. Combin. 1 (1992) 329–346. (pp. 169,
170)
[50] A. E. Brouwer & M. van Eupen, The correspondence between projective

codes and 2-weight codes, Designs, Codes and Cryptography 11 (1997)
261–266. (p. 123)
[51] A. E. Brouwer & W. H. Haemers, Structure and uniqueness of the

(81, 20, 1, 6) strongly regular graph, Discr. Math. 106/107 (1992) 77–82.
(pp. 117, 128, 188)
[52] A. E. Brouwer & W. H. Haemers, The Gewirtz graph: An exercise in the

theory of graph spectra, Europ. J. Combin. 14 (1993) 397–407. (pp. 127,
188, 189)
[53] A. E. Brouwer & W. H. Haemers, Eigenvalues and perfect matchings, Lin.

Alg. Appl. 395 (2005) 155–162. (pp. 73, 74)
[54] A. E. Brouwer & W. H. Haemers, A lower bound for the Laplacian eigen-
values of a graph—proof of a conjecture by Guo, Lin. Alg. Appl. 429
(2008) 2131–2135. (p. 49)
[55] A. E. Brouwer & J. H. Koolen, The vertex connectivity of a distance-

regular graph, Europ. J. Combin. 30 (2009) 668–673. (p. 160)
[56] A. E. Brouwer & J. H. van Lint, Strongly regular graphs and partial geome-
tries, in: Enumeration and designs, Proc. Silver Jubilee Conf. on Combi-
natorics, Waterloo 1982, eds. D. M. Jackson & S. A. Vanstone, Academic
Press, Toronto. pp. 85–122. (pp. 106, 112, 130, 142, 170)
[57] A. E. Brouwer & D. M. Mesner, The connectivity of strongly regular

graphs, Europ. J. Combin. 6 (1985) 215–216. (p. 111)
[58] A. E. Brouwer & A. Neumaier, A remark on partial linear spaces of girth

5 with an application to strongly regular graphs, Combinatorica 8 (1988)
57–61. (p. 106)
[59] A. E.pBrouwer & A. Neumaier, The graphs with spectral radius between 2
√
and 2 + 5, Lin. Alg. Appl. 114/115 (1989) 273–276. (p. 41)
[60] A. E. Brouwer & E. Spence, Cospectral graphs on 12 vertices, Electr. J.

Combin. 16 (2009) N20. (p. 181)
[61] A. E. Brouwer & H. A. Wilbrink, Block designs, in: Handbook of Incidence

Geometry (F. Buekenhout ed.), Elsevier, 1995, pp. 349–382. (p. 76)
BIBLIOGRAPHY 207
[62] A. E. Brouwer, R. M. Wilson & Qing Xiang, Cyclotomy and Strongly

Regular Graphs, J. Alg. Combin. 10 (1999) 25–28. (p. 126)
[63] R. A. Brualdi & A. J. Hoffman, On the spectral radius of (0,1) matrices,

Lin. Alg. Appl. 65 (1985) 133–146. (p. 42)
[64] N. G. de Bruijn, A combinatorial problem, Nederl. Akad. Wetensch. 49

(1946) 758–764 = Indag. Math. 8 (1946) 461–467. (p. 25)
[65] N. G. de Bruijn, Acknowledgement of priority to C. Flye sainte-Marie

on the counting of circular arrangements of 2n zeros and ones that show
each n-letter word exactly once, Report 75-WSK-06, Techn. Hogeschool
Eindhoven. (p. 25)
[66] Kurt Bryan & Tanya Leise, The $25,000,000,000 eigenvector—The linear
algebra behind Google, SIAM Review 48 (2006) 569–581. (p. 56)
[67] M. Burrow, Representation theory of finite groups, Academic Press, New

York, 1965. (p. 148)
[68] F. C. Bussemaker & D. M. Cvetković, There are exactly 13 connected,

cubic, integral graphs, Univ. Beograd Publ. Elektrotehn. Fak., Ser. Mat.
Fiz., No 544–576, (1976) 43–48. (p. 44)
[69] F. C. Bussemaker, D.M. Cvetković & J. J. Seidel, Graphs related to ex-

ceptional root systems, T.H.-Report 76-WSK-05, Eindhoven University of
Technology, 1976. (pp. 185, 186)
[70] F. C. Bussemaker, W. H. Haemers, R. Mathon & H. A. Wilbrink, A

(49,16,3,6) strongly regular graph does not exist, Europ. J. Combin. 10
(1989) 413–418. (p. 127)
[71] F. C. Bussemaker, R. A. Mathon & J. J. Seidel, Tables of two-graphs, pp.

70–112 in: Combinatorics and Graph Theory (S. B. Rao, ed.), Springer
LNM 885, 1981. Also Report 79-WSK-05, Techn. Univ. Eindhoven, 1979.
(p. 135)
[72] D. de Caen, E. R. van Dam & E. Spence, A nonregular analogue of con-

ference graphs, J. Combin. Th. (A) 88 (1999) 194–204. (p. 200)
[73] A. R. Calderbank & W. M. Kantor, The geometry of two-weight codes,

Bull. London Math. Soc. 18 (1986) 97–122. (p. 120)
[74] P. J. Cameron, Orbits of permutation groups on unordered sets. II, J.

Lond. Math. Soc., II. Ser. 23 (1981) 249–264. (p. 115)
[75] P. J. Cameron, Automorphism groups of graphs, pp. 89–127 in: Selected

Topics in Graph Theory 2, L. W. Beineke & R. J. Wilson, eds., Acad.
Press, 1983. (p. 153)
[76] P. J. Cameron, J.-M. Goethals & J. J Seidel, Strongly regular graphs having
strongly regular subconstituents, J. Algebra 55 (1978) 257–280. (pp. 117,
188)
208 BIBLIOGRAPHY
[77] P. J. Cameron, J.-M. Goethals, J. J. Seidel & E. E. Shult, Line graphs,

root systems, and elliptic geometry, J. Algebra 43 (1976) 305–327. (pp. 93,
185)
[78] A. Cayley, A theorem on trees, Quart. J. Pure Appl. Math. 23 (1889)

376–378. (p. 15)
[79] L. C. Chang, The uniqueness and nonuniqueness of the triangular associ-

ation scheme, Sci. Record 3 (1959) 604–613. (p. 110)
[80] L. C. Chang, Association schemes of partially balanced block designs with

parameters v = 28, n1 = 12, n2 = 15 and p211 = 4, Sci. Record 4 (1960)
12–18. (pp. 110, 127, 188)
[81] Gregory Cherlin, Locally finite generalized quadrangles with at most five
points per line, Discr. Math. 291 (2005) 73–79. (p. 115)
[82] A. A. Chesnokov & W. H. Haemers, Regularity and the generalized adja-

cency spectra of graphs, Lin. Alg. Appl. 416 (2006) 1033–1037. (p. 182)
[83] B. Cheyne, V. Gupta & C. Wheeler, Hamilton cycles in addition graphs,

Rose-Hulman Undergraduate Math. Journal 1 (2003) (electronic). (p. 90)
[84] H. Chuang, & G. R. Omidi, Graphs with three distinct eigenvalues and
largest eigenvalue less than 8, Lin. Alg. Appl. 430 (2009) 2053–2062.
(p. 199)
[85] F. R. K. Chung, Diameters and eigenvalues, J. Amer. Math. Soc. 2 (1989)

187–196. (pp. 70, 90)
[86] Fan R. K. Chung, Spectral Graph Theory, CBMS Lecture Notes 92, Amer.
Math. Soc., Providence, 1997. (Not cited)
[87] V. Chvátal, Tough graphs and hamiltonian circuits, Discr. Math. 5 (1973)
215–228. (p. 69)
[88] V. Chvátal & P. Erdős, A note on Hamiltonian circuits, Discr. Math. 2

(1972) 111–113. (p. 112)
[89] S. M. Cioabă, Perfect matchings, eigenvalues and expansion, C. R. Math.

Acad. Sci. Soc. R. Can. 27 (4) (2005) 101–104. (p. 74)
[90] S. M. Cioabă, D. A. Gregory & W. H. Haemers, Matchings in regular

graphs from eigenvalues, J. Combin. Th. (B) 99 (2009) 287–297. (p. 74)
[91] Sebastian M. Cioabă, Edwin R. van Dam, Jack H. Koolen & Jae-Ho
Lee, Asymptotic results on the spectral radius and the diameter of graphs,
preprint, 2008. (p. 41)
[92] W. H. Clatworthy, Partially balanced incomplete block designs with two

associate classes and two treatments per block, J. Res. Nat. Bur. Standards
54 (1955) 177–190. (p. 126)
[93] A. Clebsch, Ueber die Flächen vierter Ordnung, welche eine Doppelcurve
zweiten Grades besitzen, J. für Math. 69 (1868) 142–184. (p. 126)
BIBLIOGRAPHY 209
[94] L. Collatz & U. Sinogowitz, Spektren endlicher Grafen, Abh. Math. Sem.
Univ. Hamburg 21 (1957) 63–77. (p. 26)
[95] K. Coolsaet, J. Degraer & E. Spence, The Strongly Regular (45,12,3,3)

graphs, Electr. J. Combin. 13 (2006) R32. (p. 127)
[96] K. Coolsaet & J. Degraer, A computer assisted proof of the uniqueness of

the Perkel graph, Designs, Codes and Cryptography 34 (2005) 155–171.
(p. 190)
[97] K. Coolsaet & J. Degraer, Using algebraic properties of minimal idempo-

tents for exhaustive computer generation of association schemes, Electr.
J. Combin. 15 (2008) R30. (p. 188)
[98] R. Courant & D. Hilbert, Methoden der Mathematischen Physik, Springer

Verlag, Berlin, 1924. (p. 33)
[99] H. S. M. Coxeter, Self-dual configurations and regular graphs, Bull. Amer.

Math. Soc. 56 (1950) 413–455. (p. 126)
[100] R. Craigen & H. Kharaghani, Hadamard matrices and Hadamard de-

signs, in: Handbook of Combinatorial Designs, 2nd ed., (C.J. Colbourn
and J.H. Denitz eds.), Chapman & Hall/CRC press, Boca Raton, 2007,
Chapter V.1. (p. 141)
[101] Péter Csikvári, On a conjecture of V. Nikiforov, Discr. Math. (2009) 309

4522–4526. (p. 42)
[102] Péter Csikvári, Integral trees of arbitrarily large diameters, preprint, 2009.
(p. 85)
[103] D. M. Cvetković, Graphs and their spectra, Publ. Elektrotehn. Fak. Ser.
Mat. Fiz. 354–356 (1971) 1–50. (p. 62)
[104] D. M. Cvetković & M. Doob, Root systems, forbidden subgraphs and spec-
tral characterizations of line graphs, in: Graph Theory, Proc. Fourth Yu-
goslav Sem. Graph Theory, Novi Sad, 1983 (Cvetković, Gutman, Pisanski,
Tošić eds.), pp. 69–99. (p. 185)
[105] D. M. Cvetković, M. Doob, I. Gutman & A. Torgašev, Recent Results in

the Theory of Graph Spectra, Annals of Discr. Math. 36, North Holland,
Amsterdam, 1988. (pp. 176, 185)
[106] D. M. Cvetković, M. Doob & H. Sachs, Spectra of Graphs: Theory and

Applications, V.E.B. Deutscher Verlag der Wissenschaften, Berlin, 1979.
(Also: Academic Press, New York, 1980. Third edition, Johann Abrosius
Barth Verlag, Heidelberg-Leipzig, 1995. Third rev. enl. ed., Wiley, New
York, 1998.) (pp. 21, 176, 177, 189)
[107] D. Cvetković, M. Lepović & S. K. Simić, The maximal exceptional graphs,

J. Combin. Th. (B) 86 (2002) 347–363. (p. 66)
[108] D. Cvetković, P. Rowlinson & S. K. Simić, Eigenspaces of graphs, Cam-

bridge University Press, Cambridge, 1997. (p. 66)
210 BIBLIOGRAPHY
[109] D. Cvetković, P. Rowlinson & S. K. Simić, Graphs with least eigenvalue

−2: The star complement technique, J. Alg. Combin. 14 (2001) 5–16.
(p. 66)
[110] Dragoš Cvetković, Peter Rowlinson & Slobodan K. Simić, Spectral gener-
alizations of line graphs; on graphs with least eigenvalue −2, LMS Lecture
Note Series 314, Cambridge University Press, Cambridge, 2004. (pp. 66,
186)
[111] Dragoš Cvetković, Peter Rowlinson & Slobodan K. Simić, Signless Lapla-
cians of finite graphs, Lin. Alg. Appl. 423 (2007) 155–171. (Not cited)
[112] E. R. van Dam, Regular graphs with four eigenvalues, Lin. Alg. Appl. 226-
228 (1995) 139–162. (pp. 195, 196)
[113] E. R. van Dam, Graphs with few eigenvalues; an interplay between combi-
natorics and algebra, CentER Dissertation 20, Tilburg University, 1996.
(pp. 196, 198, 199)
[114] E. R. van Dam, Nonregular graphs with three eigenvalues, J. Combin. Th.
(B) 73 (1998) 101–118. (p. 199)
[115] E. R. van Dam, The spectral excess theorem for distance-regular graphs:
a global (over)view, Electr. J. Combin. 15 (2008) R129. (p. 163)
[116] E. R. van Dam and W. H. Haemers, Graphs with constant µ and µ, Discr.
Math. 182 (1998) 293–307. (p. 198)
[117] E. R. van Dam & W. H. Haemers, Spectral characterizations of some
distance-regular graphs, J. Alg. Combin. 15 (2002) 189–202. (pp. 175,
176, 187, 189)
[118] E. R. van Dam & W. H. Haemers, Which graphs are determined by their
spectrum?, Lin. Alg. Appl. 373 (2003) 241–272. (p. 175)
[119] E. R. van Dam, W. H. Haemers & J. H. Koolen, Cospectral graphs and the
generalized adjacency matrix, Lin. Alg. Appl. 423 (2007) 33–41. (pp. 175,
176)
[120] E. R. van Dam, W. H. Haemers, J. H. Koolen & E. Spence, Characterizing
distance-regularity of graphs by the spectrum, J. Combin. Th. (A) 113
(2006) 1805–1820. (pp. 187, 189)
[121] E. R. van Dam & J. H. Koolen, A new family of distance-regular graphs
with unbounded diameter, Invent. Math. 162 (2005) 189–193. (pp. 157,
159)
[122] E. R. van Dam & E. Spence, Small regular graphs with four eigenvalues,
Discr. Math. 189 (1998) 233–257. (pp. 195, 196)
[123] R. M. Damerell, On Moore graphs, Proc. Cambridge Philos. Soc. 74 (1973)
227–236. (p. 161)
[124] K. C. Das, A characterization of graphs which achieve the upper bound
for the largest Laplacian eigenvalue of graphs, Lin. Alg. Appl. 376 (2004)
173–186. (p. 48)
BIBLIOGRAPHY 211
[125] J. Degraer & K. Coolsaet, Classification of some strongly regular subgraphs

of the McLaughlin graph, Discr. Math. 308 (2008) 395–400. (p. 188)
[126] Ph. Delsarte, Weights of linear codes and strongly regular normed spaces,
Discr. Math. 3 (1972) 47–64. (p. 120)
[127] Ph. Delsarte, An algebraic approach to the association schemes of coding

theory, Philips Res. Rep. Suppl. 10 (1973). (pp. 147, 151, 154)
[128] Ph. Delsarte, J.-M. Goethals & J. J. Seidel, Orthogonal matrices with zero
diagonal, Part II, Canad. J. Math. 23 (1971) 816–832. (p. 139)
[129] Ph. Delsarte, J.-M. Goethals & J. J. Seidel, Bounds for systems of lines
and Jacobi polynomials, Philips Res. Rep. 30 (1975) 91*–105*. (pp. 142,
144, 145)
[130] Ph. Delsarte, J.-M. Goethals & J. J. Seidel, Spherical codes and designs,
Geom. Dedicata 6 (1977) 363–388. (p. 145)
[131] M. DeVos, L. Goddyn, B. Mohar & R. Šámal, Cayley sum graphs and
eigenvalues of (3,6)-fullerenes, J. Combin. Th. (B) 99 (2009) 358–369.
(p. 90)
[132] M. Doob, Graphs with a small number of distinct eigenvalues, Ann. N.Y.
Acad. Sci. 175 (1970) 104–110. (p. 195)
[133] M. Doob, On characterizing certain graphs with four eigenvalues by their

spectrum, Linear Algebra Appl. 3 (1970) 461–482. (p. 195)
[134] M. Doob & W. H. Haemers, The complement of the path is determined by

its spectrum, Lin. Alg. Appl. 356 (2002) 57–65. (p. 184)
[135] Joshua Ducey & Peter Sin, The elementary divisors of the incidence matri-
ces of skew lines in P G(3, p), preprint, 2010. arXiv:1001.2551v1 (p. 174)
[136] Art M. Duval & Victor Reiner, Shifted simplicial complexes are Laplacian
integral, Trans. Amer. Math. Soc. 354 (2002) 4313–4344. (p. 54)
[137] M. N. Ellingham, Basic subgraphs and graph spectra, Australasian J. Com-

bin. 8 (1993) 247–265. (p. 66)
[138] Ky Fan, On a theorem of Weyl concerning eigenvalues of linear transfor-

mations I, Proc. Nat. Acad. Sci. USA 35 (1949) 652–655. (p. 34)
[139] W. Feit & G. Higman, The non-existence of certain generalized polygons,

J. Algebra 1 (1964) 114–131. (p. 161)
[140] N. C. Fiala & W. H. Haemers, 5-Chromatic strongly regular graphs, Discr.

Math. 306 (2006) 3083–3096. (p. 113)
[141] M. Fiedler, Algebraic connectivity of graphs, Czech. Math. J. 23 (98)

(1973) 298–305. (p. 21)
[142] M. Fiedler, Eigenvectors of acyclic matrices, Czech. Math. J. 25 (1975)

607–618. (p. 81)
212 BIBLIOGRAPHY
[143] M. A. Fiol & E. Garriga, From local adjacency polynomials to locally

pseudo-distance-regular graphs, J. Combin. Th. (B) 71 (1997) 162–183.
(pp. 163, 189)
[144] M. A. Fiol, S. Gago & E. Garriga, A simple proof of the spectral excess
theorem for distance-regular graphs, Lin. Alg. Appl. 432 (2010) 2418–2422.
(p. 163)
[145] C. Flye Sainte-Marie, Solution to question nr. 48, l’Intermédiaire des
Mathématiciens 1 (1894) 107–110. (p. 25)
[146] P. W. Fowler, P. E. John & H. Sachs, (3,6)-cages, hexagonal toroidal cages,
and their spectra, Discrete Mathematical Chemistry (New Brunswick, NJ,
1998), DIMACS Ser. Discr. Math. Theoret. Comp. Sci. vol. 51, AMS,
Providence, RI, 2000, pp. 139–174. (p. 90)
[147] J. Friedman, A proof of Alon’s second eigenvalue conjecture and related
problems, Memoirs of AMS 195, 2008, 100pp, and arXiv:cs/0405020v1
(118 pages). See also: A proof of Alon’s second eigenvalue conjecture, pp.
720–724 in: STOC ’03: Proc. 35th annual ACM Symp. on Theory of
Computing, San Diego, 2003, ACM, New York, 2003. (p. 67)
[148] H. Fujii & A. Katsuda, Isospectral graphs and isoperimetric constants,
Discr. Math. 207 (1999) 33–52. (p. 176)
[149] F. R. Gantmacher, Applications of the Theory of Matrices, Interscience,
New York, 1959. (pp. 30, 35)
[150] A. L. Gavrilyuk & A. A. Makhnev, On Krein Graphs without Triangles,
Dokl. Akad. Nauk 403 (2005) 727–730 (Russian) / Doklady Mathematics
72 (2005) 591–594 (English). (p. 105)
[151] S. Geršgorin, Über die Abgrenzung der Eigenwerte einer Matrix, Izv. Akad.
Nauk. SSSR Otd. Fiz.-Mat. Nauk 7 (1931) 749–754. (p. 36)
[152] A. Gewirtz, Graphs with maximal even girth, Canad. J. Math. 21 (1969)
915–934. (pp. 127, 129, 188)
[153] A. Gewirtz, The uniqueness of g(2, 2, 10, 56), Trans. New York Acad. Sci.
31 (1969) 656–675. (p. 127)
[154] C. D. Godsil, Algebraic Combinatorics, Chapman and Hall, New York,
1993. (pp. 31, 176, 177)
[155] C. D. Godsil & B. D. McKay, Some computational results on the spectra
of graphs, in: Combinatorial Mathematics IV (L.R.A. Casse and W.D.
Wallis eds.; Proc. fourth Australian conf. on Combin. Math., Adelaide),
Lecture Notes in Mathematics 560, Springer-Verlag, Berlin 1976, pp. 73–
92. (pp. 177, 180)
[156] C. D. Godsil & B. D. McKay, Constructing cospectral graphs, Aequat.
Math. 25 (1982) 257–268. (pp. 23, 178, 180, 181)
[157] C. D. Godsil, Spectra of trees, Annals of Discr. Math. 20 (1984) 151–159.
(p. 80)
BIBLIOGRAPHY 213
[158] C. D. Godsil & G. Royle, Algebraic Graph Theory, Springer-Verlag, New

York, 2001. (p. 113)
[159] J.-M. Goethals & J. J. Seidel, Orthogonal matrices with zero diagonal,
Canad. J. Math. 19 (1967) 1001–1010. (p. 140)
[160] J.-M. Goethals & J. J. Seidel, Strongly regular graphs derived from com-
binatorial designs, Canad. J. Math. 22 (1970) 597–614. (pp. 130, 131,
142)
[161] J.-M. Goethals & J. J. Seidel, The regular two-graph on 276 vertices, Discr.
Math. 12 (1975) 143–158. (pp. 138, 139, 188)
[162] R. L. Graham & H. O. Pollak, On the addressing problem for loop switch-
ing, Bell System Tech. J. 50 (1971) 2495–2519. (p. 20)
[163] M. Grassl, On SIC-POVMs and MUBs in dimension 6, quant-ph/

0406175. (p. 143)
[164] B. J. Green, Counting sets with small sumset, and the clique number of
random Cayley graphs, Combinatorica 25 (2005) 307–326. (p. 90)
[165] R. Grone, Eigenvalues and degree sequences of graphs, Lin. Multilin.

Alg. 39 (1995) 133–136. (p. 48)
[166] R. Grone & R. Merris, The Laplacian spectrum of a graph II, SIAM J.
Discr. Math. 7 (1994) 221–229. (pp. 48, 50)
[167] Robert Grone, Russell Merris & William Watkins, Laplacian unimodular
equivalence of graphs, pp. 175–180 in: Combinatorial and Graph-Theoretic
Problems in Linear Algebra (R. Brualdi, S. Friedland & V. Klee, eds.),
IMA Volumes in Mathematics and Its Applications 50, Springer, NY,
1993. (p. 172)
[168] D. Grynkiewicz, V. F. Lem & O. Serra, The connectivity of addition Cayley

graphs, Electr. Notes in Discr. Math. 29 (2007) 135–139. (p. 90)
[169] Stephen Guattery & Gary L. Miller, On the performance of the spectral
graph partitioning methods, pp. 233–242 in: Proc. 2nd ACM-SIAM Sym-
posium on Discrete Algorithms, 1995. (p. 58)
[170] Ji-Ming Guo, A new upper bound for the Laplacian spectral radius of
graphs, Lin. Alg. Appl. 400 (2005) 61–66. (p. 48)
[171] Ji-Ming Guo, On the third largest Laplacian eigenvalue of a graph, Lin.
Multilin. Alg. 55 (2007) 93–102. (p. 49)
[172] W. H. Haemers, An upper bound for the Shannon capacity of a graph,

Colloq. Math. Soc. János Bolyai Vol 25, Algebraic Methods in Graph
Theory (1978) 267–272. (p. 65)
[173] W. H. Haemers, On some problems of Lovász concerning the Shannon

capacity of graphs, IEEE Trans. Inf. Th. 25 (1979) 231–232. (p. 65)
214 BIBLIOGRAPHY
[174] W. H. Haemers, Eigenvalue methods, in: Packing and covering in combi-

natorics (A. Schrijver ed.), Math. Centre Tract 106, Mathematical Centre,
Amsterdam, 1979, pp. 15–38. (p. 63)
[175] W. H. Haemers, Eigenvalue techniques in design and graph theory, Reidel,

Dordrecht, 1980. Thesis (T.H. Eindhoven, 1979) = Math. Centr. Tract 121
(Amsterdam, 1980). (pp. 33, 63, 112)
[176] W. H. Haemers, There exists no (76,21,2,7) strongly regular graph, pp.

175–176 in: Finite Geometry and Combinatorics, F. De Clerck et al. (eds.),
LMS Lecture Notes Series 191, Cambridge University Press, 1993. (p. 128)
[177] W. H. Haemers, Interlacing eigenvalues and graphs, Lin. Alg. Appl. 226–
228 (1995) 593–616. (p. 33)
[178] W. H. Haemers, Distance-regularity and the spectrum of graphs, Lin. Alg.
Appl. 236 (1996) 265–278. (p. 189)
[179] W. H. Haemers, Strongly regular graphs with maximal energy, Lin. Alg.
Appl. 429 (2008) 2719–2723. (p. 142)
[180] W. H. Haemers, A. Mohammadian & B. Tayfeh-Rezaie, On the sum of

laplacian eigenvalues of graphs, preprint, 2008. (p. 50)
[181] W. H. Haemers & C. Roos, An inequality for generalized hexagons, Geom.

Dedicata 10 (1981) 219–222. (p. 161)
[182] W. H. Haemers & M. S. Shrikhande, Some remarks on subdesigns of sym-

metric designs, J. Stat. Plann. Infer. 3 (1979) 361–366. (p. 76)
[183] W. H. Haemers & E. Spence, Graphs cospectral with distance-regular

graphs, Lin. Multilin. Alg. 39 (1995) 91–107. (pp. 158, 168, 176, 187,
188, 189)
[184] W. H. Haemers & E. Spence, The Pseudo-Geometric Graphs for Gener-
alised Quadrangles of Order (3,t), Europ. J. Combin. 22 (2001) 839–845.
(p. 128)
[185] W. H. Haemers & E. Spence, Enumeration of cospectral graphs, Europ. J.

Combin. 25 (2004) 199–211. (p. 180)
[186] W. H. Haemers & Q. Xiang, Strongly regular graphs with parameters

(4m4 , 2m4 + m2 , m4 + m2 , m4 + m2 ) exist for all m > 1, Europ. J. Com-
bin. 31 (2010) 1553–1559. (p. 142)
[187] L. Halbeisen & N. Hungerbühler, Generation of isospectral graphs, J.

Graph Theory 31 (1999) 255–265. (pp. 176, 179)
[188] N. Hamada & T. Helleseth, A characterization of some {3v2 + v3 , 3v1 +

v2 ; 3, 3}-minihypers and some [15, 4, 9; 3]-codes with B2 = 0, J. Stat.
Plann. Infer. 56 (1996) 129–146. (p. 129)
[189] C. Helmberg, B. Mohar, S. Poljak & F. Rendl, A spectral approach to

bandwidth and separator problems in graphs, Proceedings third IPCO conf.
(G. Rinaldi and L. Wolsey, eds.) (1993) 183–194. (p. 72)
BIBLIOGRAPHY 215
[190] J. van den Heuvel, Hamilton cycles and eigenvalues of graphs, Lin. Alg.
Appl. 226–228 (1995) 723–730. (p. 70)
[191] D. G. Higman, Partial geometries, generalized quadrangles and strongly

regular graphs, in: Atti Convegno di Geometria, Combinatoria e sue Ap-
plicazioni, Perugia 1970, (A. Berlotti, ed.) (1971) pp. 263–293. (p. 161)
[192] D. G. Higman & C. Sims, A simple group of order 44,352,000, Math. Z.

105 (1968) 110–113. (p. 129)
[193] R. Hill, Caps and Groups, pp. 389–394 in: Coll. Intern. Teorie Combin.
Acc. Naz. Lincei, Roma 1973, Atti dei convegni Lincei 17, Rome, 1976.
(p. 124)
[194] A. J. Hoffman, On the polynomial of a graph, Amer. Math. Monthly 70

(1963) 30–36. (pp. 22, 43)
[195] A. J. Hoffman, On eigenvalues and colourings of graphs, in: Graph theory

and its applications, (B. Harris, ed.), Acad. Press, New York, 1970, pp.
79–91. (p. 62)
[196] A. J. Hoffman & R. R. Singleton, On Moore graphs with diameter 2 and

3, IBM J. Res. Develop. 4 (1960) 497–504. (pp. 104, 127)
[197] A. J. Hoffman & J. H. Smith, On the spectral radii of topologically equiv-

alent graphs, pp. 273–281 in: Recent Advances in Graph Theory, Proc.
Symp. Prague June 1974, Academia Praha, Prague, 1975. (p. 41)
[198] S. G. Hoggar, Two quaternionic 4-polytopes, pp. 219–230 in: The Geo-
metric Vein, Coxeter Festschrift, C. Davis (ed.), Springer, 1981. (p. 143)
[199] S. G. Hoggar, 64 Lines from a Quaternionic Polytope, Geom. Dedicata

69 (1998) 287–289. (p. 143)
[200] R. A. Horn & C. R. Johnson, Matrix Analysis, Cambridge University

Press, 1985. (p. 30)
[201] T. Huang & C. Liu, Spectral characterization of some generalized odd

graphs, Graphs Comb. 15 (1999) 195–209. (p. 189)
[202] X. Hubaut, Strongly regular graphs, Discr. Math. 13 (1975) 357–381.

(p. 106)
[203] James E. Humphreys, Reflection groups and Coxeter groups, Cambridge

Univ. Press, Cambridge, 1990. (p. 41)
[204] Y. J. Ionin & M. S. Shrikhande, Combinatorics of symmetric designs,

Cambridge University Press, 2006. (p. 142)
[205] I. Martin Isaacs, Character Theory of Finite Groups, Academic Press,

1976. (p. 165)
[206] A. A. Ivanov & S. V. Shpectorov, The association schemes of dual polar

spaces of type 2 A2d−1 (pf ) are characterized by their parameters if d > 3,
Lin. Alg. Appl. 114 (1989) 113–139. (p. 117)
216 BIBLIOGRAPHY
[207] P. S. Jackson & P. Rowlinson, On graphs with complete bipartite star

complements, Lin. Alg. Appl. 298 (1999) 9–20. (p. 66)
[208] C. R. Johnson & M. Newman, A note on cospectral graphs, J. Combin.

Th. (B) 28 (1980) 96–103. (p. 175)
[209] L. K. Jørgensen & M. Klin, Switching of edges in strongly regular graphs.

I. A family of partial difference sets on 100 vertices, Electr. J. Combin. 10
(2003) R17. (pp. 129, 132, 142)
[210] D. Jungnickel, On subdesigns of symmetric designs, Math. Z. 181 (1982)

383–393. (p. 76)
[211] Taher H. Haveliwala and Sepandar D. Kamvar, The Second Eigenvalue of

the Google Matrix, Technical Report 2003-20, Stanford University, 2003.
(p. 56)
[212] W. M. Kantor & R. A. Liebler, The rank three permutation representations

of the finite classical groups, Trans. Amer. Math. Soc. 271 (1982) 1–71.
(p. 106)
[213] P. Kaski & P. R. J. Östergård, There are exactly five biplanes with k = 11,
J. Combin. Designs 16 (2008) 117–127. (p. 105)
[214] J. B. Kelly, A characteristic property of quadratic residues, Proc. Amer.

Math. Soc. 5 (1954) 38–46. (p. 104)
[215] A. K. Kelmans, On graphs with randomly deleted edges, Acta Math. Acad.
Sci. Hung. 37 (1981) 77–88. (p. 42)
[216] G. Kirchhoff, Über die Auflösung der Gleichungen, auf welche man bei der
untersuchung der linearen verteilung galvanischer Ströme geführt wird,
Ann. Phys. Chem. 72 (1847) 497–508. (p. 15)
[217] M. Klin & M. Muzychuk, On graphs with three eigenvalues, Discr.

Math. 189 (1998) 191–207. (p. 199)
[218] Yehuda Koren, On spectral graph drawing, pp. 496–508 in: COCOON
2003: Computing and Combinatorics, Springer LN Comp. Sci. 2697, 2003.
(p. 56)
[219] M. Krivelevich & B. Sudakov, Sparse pseudo-random graphs are Hamilto-

nian, J. Graph Th. 42 (2003) 17–33. (p. 70)
[220] C. L. M. de Lange, Some new cyclotomic strongly regular graphs, J. Alg.

Combin. 4 (1995) 329–330. (p. 126)
[221] P. W. H. Lemmens & J. J. Seidel, Equiangular lines, J. Algebra 24 (1973)

494–512. (pp. 40, 142, 143, 144)
[222] Jiong-Sheng Li & Yong-Liang Pan, A note on the second largest eigenvalue
of the Laplacian matrix of a graph, Lin. Multilin. Alg. 48 (2000) 117–121.
(p. 49)
BIBLIOGRAPHY 217
[223] Jiong-Sheng Li & Yong-Liang Pan, De Caen’s inequality and bounds on

the largest Laplacian eigenvalue of a graph, Lin. Alg. Appl. 328 (2001)
153–160. (p. 48)
[224] Jiong-Sheng Li & Xiao-Dong Zhang, On the Laplacian eigenvalues of a

graph, Lin. Alg. Appl. 285 (1998) 305–307. (p. 48)
[225] M. W. Liebeck, The affine permutation groups of rank three, Proc. London
Math. Soc. (3) 54 (1987) 477–516. (p. 106)
[226] M. W. Liebeck & J. Saxl, The finite primitive permutation groups of rank
three, Bull. London Math. Soc. 18 (1986) 165–172. (p. 106)
[227] J. H. van Lint, Introduction to coding theory, Graduate Texts in Math.

86, Springer, New York, 1982. (p. 120)
[228] J. H. van Lint & A. Schrijver, Constructions of strongly regular graphs,

two-weight codes and partial geometries by finite fields, Combinatorica 1
(1981) 63–73. (pp. 118, 130)
[229] J. H. van Lint & J. J. Seidel, Equilateral point sets in elliptic geometry,
Indag. Math. 28 (1966) 335–348. (pp. 142, 144)
[230] P. Lisoněk, New maximal two-distance sets, J. Combin. Th. (A) 77 (1997)
318–338. (p. 145)
[231] Huiqing Liu, Mei Lu & Feng Tian, On the Laplacian spectral radius of a
graph, Lin. Alg. Appl. 376 (2004) 135–141. (p. 48)
[232] L. Lovász, On the Shannon capacity of a graph, IEEE Trans. Inf. Th. 25
(1979) 1–7. (p. 65)
[233] A. Lubotzky, Cayley graphs: eigenvalues, expanders and random walks,

pp. 155–189 in: Surveys in combinatorics, 1995 (Stirling), London Math.
Soc. Lecture Note Ser., 218, Cambridge Univ. Press, Cambridge, 1995.
(p. 176)
[234] A. Lubotzky, R. Phillips & P. Sarnak, Ramanujan Graphs, Combinatorica

8 (1988) 261–277. (p. 67)
[235] Ulrike von Luxburg, A Tutorial on Spectral Clustering, Statistics and

Computing 17 (2007) 395–416. (p. 58)
[236] F. J. MacWilliams & N. J. A. Sloane, The Theory of Error-Correcting

Codes, North Holland Publ. Co., Amsterdam, 1977. (pp. 120, 169)
[237] C. L. Mallows & N. J. A. Sloane, Two-graphs, switching classes and Euler

graphs are equal in number, SIAM J. Appl. Math. 28 (1975) 876–880.
(p. 135)
[238] M. Marcus & H. Minc, A survey of matrix theory and matrix inequalities,
Allyn and Bacon, Boston, 1964. (p. 30)
[239] M. Marcus & H. Minc, Introduction to linear algebra, Macmillan, New

York, 1965. (p. 148)
218 BIBLIOGRAPHY
[240] G. A. Margulis, Explicit group-theoretical constructions of combinatorial

schemes and their application to the design of expanders and concentra-
tors, Probl. Peredachi Inf. 24 (1988) 51–60. English: Problems of Infor-
mation Transmission 24 (1988) 39–46. (p. 67)
[241] R. A. Mathon, Symmetric conference matrices of order pq 2 + 1, Canad. J.

Math. 30 (1978) 321–331. (p. 127)
[242] R. J. McEliece & H. Rumsey, jr., Euler products, cyclotomy and coding,
J. Number Th. 4 (1972) 302–311. (p. 125)
[243] B. D. McKay, On the spectral characterisation of trees, Ars Combinatoria

3 (1979) 219–232. (p. 177)
[244] B. D. McKay & E. Spence, Classification of regular two-graphs on 36 and

38 vertices, Australas. J. Comb. 24 (2001) 293–300. (pp. 127, 174, 188)
[245] R. Merris, Large families of Laplacian isospectral graphs, Lin. Multilin.

Alg. 43 (1997) 201–205. (p. 176)
[246] R. Merris, A note on Laplacian graph eigenvalues, Lin. Alg. Appl. 285
(1998) 33–35. (p. 48)
[247] B. Mohar, A domain monotonicity theorem for graphs and Hamiltonicity,

Discr. Appl. Math. 36 (1992) 169–177. (p. 70)
[248] Oleg R. Musin, Spherical two-distance sets, arXiv:0801.3706, 2008.

(p. 145)
[249] A. Neumaier, Strongly regular graphs with smallest eigenvalue −m, Archiv
Math. 33 (1979) 392–400. (pp. 106, 109)
[250] A. Neumaier, New inequalities for the parameters of an association

scheme, pp. 365–367 in: Combinatorics and Graph Theory, Proc. Cal-
cutta 1980, Lecture Notes in Math. Vol 885 (S.B. Rao, ed.), Springer,
Berlin. (p. 152)
[251] A. Neumaier, The second largest eigenvalue of a tree, Lin. Alg. Appl. 46
(1982) 9–25. (p. 84)
[252] A. Neumaier, Completely regular two-graphs, Arch. Math. 38 (1982) 378–

384. (pp. 138, 139)
[253] R. E. A. C. Paley, On orthogonal matrices, J. Math. Phys. Mass. Inst.

Technol. 12 (1933) 311–320. (p. 140)
[254] D. V. Pasechnik, The triangular extensions of a generalized quadrangle of

order (3,3), Simon Stevin 2 (1995) 509-518. (p. 139)
[255] A. J. L. Paulus, Conference matrices and graphs of order 26, Techni-

sche Hogeschool Eindhoven, report WSK 73/06, Eindhoven, 1983, 89 pp.
(p. 126)
[256] S. E. Payne, An inequality for generalized quadrangles, Proc. Amer. Math.

Soc. 71 (1978) 147–152. (p. 33)
BIBLIOGRAPHY 219
[257] M. J. P. Peeters, Ranks and Structure of Graphs, PhD thesis, Tilburg

University, 1995. (p. 170)
[258] René Peeters, On the p-ranks of the adjacency matrices of distance-regular
graphs, J. Alg. Combin. 15 (2002) 127–149. (p. 168)
[259] J. Petersen, Sur le théorème de Tait, L’Intermédiaire des Mathématiciens
5 (1898) 225–227. (p. 126)
[260] Tomaž Pisanski & John Shawe-Taylor, Characterizing graph drawing with
eigenvectors, J. Chem. Inf. Comp. Sci. 40 (2000) 567–571. (p. 57)
[261] L. Pyber, Large connected strongly regular graphs are Hamiltonian,
preprint, 2009. (p. 70)
[262] Gregory T. Quenell, Eigenvalue comparisons in graph theory, Pacif. J.
Math. 176 (1996) 443–461. (p. 67)
[263] P. Rowlinson, On the maximal index of graphs with a prescribed number
of edges, Lin. Alg. Appl. 110 (1988) 43–53. (p. 42)
[264] P. Rowlinson, The characteristic polynomials of modified graphs, Discr.
Appl. Math. 67 (1996) 209–219. (p. 176)
[265] B. L. Schader, On tournament matrices, Lin. Alg. Appl. 162–164 (1992)
335–368. (p. 25)
[266] I. Schur, Über eine Klasse von Mittelbildungen mit Anwendungen auf die
Determinantentheorie, Sitzungsber. Berliner Math. Ges. 22 (1923) 9–20.
(p. 33)
[267] A. J. Schwenk, Almost all trees are cospectral, in: New directions in the
theory of graphs (F. Harary ed.), Academic Press, New York 1973, pp.
275–307. (p. 177)
[268] A. J. Schwenk, Exactly thirteen connected cubic graphs have integral spec-
tra, pp. 516–533 in: Theory and Applications of graphs (Proc. Kalamazoo
1976), ed. Y. Alavi & D. Lick, Springer, 1978. (p. 44)
[269] A. J. Schwenk & O. P. Lossers, Solution to advanced problem #6434,
Amer. Math. Monthly 94 (1987) 885–886. (p. 20)
[270] L. L. Scott Jr, A condition on Higman’s parameters, Notices Amer. Math.
Soc. 20, A-97 (1973) 701-20-45. (pp. 105, 151)
[271] J. Seberry Wallis, Hadamard matrices, Part 4 in: W. D. Wallis, A. P.
Street & J. Seberry Wallis, Combinatorics: Room squares, Sum-free sets,
Hadamard matrices, Springer LNM 292, Berlin etc., 1972. (pp. 141, 142)
[272] J. Seberry & M. Yamada, Hadamard matrices, sequences, and block de-
signs, pp. 431–560 in: Contemporary design theory, J. H. Dinitz & D. R.
Stinson (eds.), Wiley, New York, 1992. (p. 141)
[273] J. J. Seidel, Strongly regular graphs of L2 -type and of triangular type, Kon.
Nederl. Akad. Wetensch. (A) 70 (1967) 188–196 = Indag. Math. 29 (1967)
188–196. (p. 139)
220 BIBLIOGRAPHY
[274] J. J. Seidel, Strongly regular graphs with (−1, 1, 0) adjacent matrix having
eigenvalue 3, Lin. Alg. Appl. 1 (1968) 281–298. (pp. 109, 126, 134, 188)
[275] J. J. Seidel, Graphs and two-graphs, in: Proc. 5th Southeast. Conf. Comb.,
Graph Th., Comp. 1974 (F. Hoffman et al., eds.), Utilitas Mathematica
Pub., Winnipeg, 1974, pp. 125–143. (p. 188)
[276] J. J. Seidel, Strongly regular graphs, in: Surveys in Combinatorics, Proc.

7th Bitish Combinatorial Conf., LMSLN ser. Vol 38, ed. B. Bollobás,
Cambridge Univ. Press, Cambridge, 1979, pp.157–180. (p. 151)
[277] J. J. Seidel, Discrete Non-Euclidean Geometry, pp. 843–920 in: Handbook

of Incidence Geometry, F. Buekenhout, ed., NHPC, 1995. (p. 144)
[278] J. J. Seidel & S. V. Tsaranov, Two-graphs, related groups, and root sys-
tems, Bull. Soc. Math. Belgique 42 (1990) 695–711. (p. 135)
[279] E. Seneta, Non-negative matrices and Markov chains, Springer Series in

Statistics, Springer, New York, 1981. (p. 30)
[280] Á. Seress, Large families of cospectral graphs, Designs, Codes and Cryp-
tography 21 (2000) 205–208. (p. 176)
[281] J.-P. Serre, Répartition asymptotique des valeurs propres de l’opérateur de

Hecke Tp , J. Amer. Math. Soc. 10 (1997) 75–102. (p. 67)
[282] C. E. Shannon, The zero-error capacity of a noisy channel, IRE Trans.

Inform. Th. vol. IT-2 (1956) 8–19. (p. 64)
[283] S. S. Shrikhande, The uniqueness of the L2 association scheme, Ann.

Math. Statist. 30 (1959) 781–798. (pp. 110, 126, 188)
[284] Peter Sin, The p-rank of the incidence matrix of intersecting linear sub-
spaces, Des. Codes Cryptogr. 31 (2004) 213–220. (p. 174)
[285] J. H. Smith, Some properties of the spectrum of a graph, pp. 403–406

in: Combinatorial Structures and their applications (Proc. Conf. Calgary,
1969), R. Guy et al. (eds.), Gordon and Breach, New York, 1970. (p. 40)
[286] E. Spence, The Strongly Regular (40,12,2,4) Graphs, Electr. J. Combin. 7

2000, R22. (p. 127)
[287] E. Spence, http://www.maths.gla.ac.uk/~es/cospecL/cospecL.html

and .../cospecSignlessL/cospecSignlessL.html (p. 181)
[288] Daniel A. Spielman & Shang-Hua Teng, Spectral partitioning works: pla-
nar graphs and finite element meshes, pp. 96–105 in: Proc. 37th Sympo-
sium on Foundations of Computer Science (Burlington, 1996), IEEE 1996.
(p. 58)
[289] Toshikazu Sunada, Riemannian coverings and isospectral manifolds, Ann.

Math. 121 (1985) 169–186. (p. 179)
BIBLIOGRAPHY 221
[290] J. J. Sylvester, A demonstration of the theorem that every homogeneous

quadratic polynomial is reducible by real orthogonal substitutions to the
form of a sum of positive and negative squares, Philos. Mag. 4 (1852)
138–142. (Reprinted in: Coll. Math. Papers 1, Cambridge Univ. Press,
1904, pp. 378–381.) (p. 35)
[291] G. Szegö, Orthogonal Polynomials, AMS Colloq. Publ. 23, 1939 (1st ed.),
1975 (4th ed.). (p. 154)
[292] R. M. Tanner, Explicit concentrators from generalized n-gons, SIAM J.
Algebr. Discr. Meth. 5 (1984) 287–293. (p. 68)
[293] D. E. Taylor, Some topics in the theory of finite groups, Ph.D. Thesis,
Univ. of Oxford, 1971. (p. 137)
[294] D. E. Taylor, Graphs and block designs associated with the three-dimen-
sional unitary groups, pp. 128–131 in: Combinatorial Math., Proc. 2nd
Austr. Conf., D. A. Holton, ed., Springer LNM 403, Berlin 1974. (p. 130)
[295] D. E. Taylor, Regular 2-graphs, Proc. London Math. Soc. 35 (1977) 257–
274. (pp. 130, 138)
[296] W. T. Tutte, The factorizations of linear graphs, J. London Math. Soc.
22 (1947) 107–111. (p. 73)
[297] W. T. Tutte, All the king’s horses. A guide to reconstruction, pp. 15–33
in: Graph Theory and Related Topics (Proc. Conf. Waterloo, 1977), J. A.
Bondy & U. S. R. Murty (eds.), Academic Press, New York, 1979. (p. 24)
[298] H. Van Maldeghem, Generalized Polygons, Birkhäuser, Basel, 1998.
(p. 161)
[299] R. S. Varga, Matrix iterative analysis, Prentice-Hall, Englewood Cliffs,
New York, 1962. (p. 30)
[300] R. S. Varga, Geršgorin and his circles, Springer, 2004. (p. 36)
[301] Wei Wang & Cheng-xian Xu, A sufficient condition for a family of graphs
being determined by their generalized spectra, Europ. J. Combin. 27 (2006)
826–840. (p. 192)
[302] Wei Wang & Cheng-xian Xu, Note: On the generalized spectral charac-
terization of graphs having an isolated vertex, Lin. Alg. Appl. 425 (2007)
210–215. (p. 193)
[303] Y. Wang, Y. Fan & Y. Tan, On graphs with three distinct Laplacian eigen-
values, Appl. Math. J. Chinese Univ. Ser. B 22 (2007) 478–484. (p. 198)
[304] Mamoru Watanabe, Note on integral trees, Math. Rep. Toyama Univ. 2
(1979) 95–100. (p. 85)
[305] H. Whitney, Congruent graphs and the connectivity of graphs, Amer. J.
Math. 54 (1932) 150–167. (p. 187)
[306] H. A. Wilbrink & A. E. Brouwer, A (57, 14, 1) strongly regular graph does
not exist, Indag. Math. 45 (1983) 117–121. (p. 127)
222 BIBLIOGRAPHY
[307] H. S. Wilf, The eigenvalues of a graph and its chromatic number, J. Lon-
don Math. Soc. 42 (1967) 330–332. (p. 62)
[308] R.√Woo & A. Neumaier, On graphs whose spectral radius is bounded by

3
2 2, Graphs Comb. 23 (2007) 713–726. (p. 41)
[309] Chen Yan, Properties of spectra of graphs and line graphs, Appl. Math. J.
Chinese Univ. Ser. B 17 (2002) 371–376. (p. 48)
[310] Gerhard Zauner, Quantendesigns, Grundzüge einer nichtkommutativen

Designtheorie, Ph.D. thesis, Vienna, 1999. (p. 143)
[311] F. Zhang, Matrix Theory, Springer-Verlag, New York, 1999. (p. 34)
[312] Xiao-Dong Zhang, Two sharp upper bounds for the Laplacian eigenvalues,
Lin. Alg. Appl. 376 (2004) 207–213. (p. 48)
[313] Xiao-Dong Zhang & Rong Luo, The spectral radius of triangle-free graphs,
Australas. J. Comb. 26 (2002) 33–39. (p. 47)
AUTHOR INDEX 223
Author Index
Ajtai, 68 Erdős, 112

Alon, 64–67, 71 van Eupen, 123
Aschbacher, 153
Ayoobi, 200 Fan, 34
Feit, 76, 161
Bagchi, 106 Fiedler, 21, 81
Bai, 50 Fiol, 162, 163, 189
Bang, 160 Foulser, 106
Bannai, 160, 161 Friedman, 67
Belevitch, 140
Berman, 30 Gago, 163
Bigalke, 112 Gantmacher, 30, 35
Blokhuis, 139, 145 Garriga, 162, 163, 189
Boppana, 66 Gavrilyuk, 105
Bose, 109, 147, 148 Geršgorin, 36
Brauer, 165 Gerzon, 143
Bridges, 199 Godsil, 31, 176–178, 180
Brouwer, 41, 49, 85, 106, 115, 123, Goethals, 93, 117, 138, 144, 185
161, 189 Graham, 20
Brualdi, 42 Gregory, 74
Burrow, 148 Grone, 48, 50, 172
Bussemaker, 44, 186 Guo, 49
Haemers, 33, 49, 65, 74, 76, 111,

Calderbank, 120
139, 161, 175, 176, 180, 182
Cameron, 93, 115, 117, 153, 185
Hasse, 125
Cauchy, 33
Helmberg, 72
Cayley, 15
Higman, 33, 137, 153, 161
Chang, 110
Hilbert, 33
Cherlin, 115
Hill, 124
Chesnokov, 182
Hoffman, 22, 41–43, 62, 104
Chuang, 199
Horn, 30
Chung, 70, 90
Huang, 189
Chvátal, 69, 112
Hubaut, 106
Cioabă, 41, 74
Cohen, 161, 189 Isaacs, 165
Courant, 33 Ito, 160, 161
Csikvári, 42, 85 Ivanov, 117
Cvetković, 21, 44, 62, 185, 186
Johnson, 30, 175
van Dam, 41, 120, 157, 159, 175, Jung, 112
176, 195, 196, 199, 200 Jungnickel, 76
Damerell, 161
Davenport, 125 Kantor, 106, 115, 120
Delsarte, 120, 144, 147, 151, 154 Kaski, 105
Doob, 21, 185, 195 Kelly, 24
Dubickas, 160 Kelmans, 42
Duval, 54 Kirchhoff, 15
224 BIBLIOGRAPHY
Klin, 199 Quenell, 67

Komlós, 68
Koolen, 41, 157, 159, 160, 175, 176 Reiner, 54
Krivelevich, 70 Rendl, 72
Roos, 161
de Lange, 126 Rowlinson, 42
Lee, 41 Rumsey, 125
Lemmens, 40, 144
Liebeck, 106 Sachs, 21
Liebler, 106 Sarnak, 67
van Lint, 106, 118, 120, 144 Saxl, 106
Lisoněk, 145 Schrijver, 118
Liu, 189 Schur, 33
Lovász, 65 Schwenk, 44, 176, 177
Lubotzky, 67 Scott, 105, 151
Luo, 47 Seidel, 40, 93, 106, 109, 117, 134,
135, 138, 142, 144, 151, 152,
MacWilliams, 120, 169 185, 186
Makhnev, 105 Seneta, 30
Marcus, 30, 148 Seress, 176
Margulis, 67 Serre, 66, 67
McEliece, 125 Seymour, 64
McKay, 176–178, 180 Shannon, 64
Mena, 199 Shawe-Taylor, 57
Merris, 50, 172 Shimamoto, 147
Mesner, 148 Shpectorov, 117
Milman, 71 Shrikhande, 76, 110
Minc, 30, 148 Shult, 93, 185
Mohar, 72 Sims, 33
Moulton, 160 Sin, 174
Musin, 145 Singleton, 104
Muzychuk, 199 Sloane, 120, 169
Smith, 40, 41
Neumaier, 41, 84, 106, 139, 152, 161, Spence, 180, 195, 196, 200
189 Stickelberger, 125
Newman, 175 Sudakov, 70
Sunada, 176
Omidi, 199, 200 Sylvester, 35
Östergård, 105 Szegö, 154
Page, 55 Szemerédi, 68
Paley, 140 Tanner, 68
Pasechnik, 139 Tayfeh-Rezaie, 200
Payne, 33 Taylor, 137
Peeters, 168 Tsaranov, 135
Phillips, 67 Tutte, 73
Pisanski, 57
Plemmons, 30 Ulam, 24
Poljak, 72
Pollak, 20 Van Maldeghem, 161
Pyber, 70 Varga, 30, 36
AUTHOR INDEX 225
Wang, 192, 193

Watanabe, 85
Watkins, 172
Whitney, 187
Wilbrink, 76
Wilf, 62
Witsenhausen, 20
Woo, 41
Xu, 192, 193
Yan, 48
Zhang, 47
226 BIBLIOGRAPHY
Subject Index
absolute bound, 106, 143, 152 cone, 26, 199

absolute point, 77 conference graphs, 104, 140
addition Cayley graphs, 90 conference matrix, 139
adjacency matrix, 11 conjunction, 19
algebraic connectivity, 21 core, 140
antipodal, 161 cospectral, 22, 23, 26, 45, 158, 175–
associated, 134 183
association scheme, 147 Courant-Weyl inequalities, 34
automorphism, 20 cover, 88
covering, 88
bandwidth, 72 Cvetković bound, 62
Bannai-Ito conjecture, 160
Berlekamp-van Lint-Seidel graph, 123, de Bruijn cycle, 25
131 de Bruijn graph, 25
biaffine plane, 191 Desargues graph, 45
biclique, 71 descendant, 135
Biggs-Smith graph, 160 determinant, 95
bipartite double, 19 diagonally dominant, 36
bipartite graph, 15, 43 diameter, 14, 70
biplane, 107 dimension, 53
block graph, 108 directed incidence matrix, 11
Bose-Mesner algebra, 148 discriminant, 95
Brouwer-Haemers graph, 117 distance-regular, 157
dominance order, 54
cap, 132 downshifting, 54
Cauchy-Binet formula, 15 DS graph, 175
Cayley graph, 19, 87 dual, 120, 123
Cayley sum graph, 90 dual lattice, 94
Chang graphs, 109, 110, 172 Dynkin diagram, 41, 96
characteristic matrix, 30
characteristic polynomial, 12 edge expansion, 69
chromatic number, 62 effective length, 120
3-claw, 93 eigenvalue components, 81
Clebsch graph, 105, 110 elementary divisors, 171
clique, 61, 138 endpath, 41
clique number, 61 energy, 57
coalescence, 177 equiangular tight frame, 143
cocktailparty graph, 93 equitable, 30
coclique, 61 equivalent code, 120
cofactor, 15 even lattice, 94
coherent, 138 exceptional graph, 99
coherent triple, 134 expander, 68, 90
collinearity graph, 108, 178 expander mixing lemma, 68
cometric association scheme, 154 extended fundamental system, 96
companion matrix, 192
complement, 14, 102, 135 feasible, 126
completely regular two-graph, 138 few eigenvalues, 195
SUBJECT INDEX 227
firm generalized n-gon, 114 imprimitive srg, 103

flag, 75 improper strongly regular graph, 133
flag graph, 190 incidence graph, 75
folded 5-cube, 105 incidence matrix, 11
Fowler conjecture, 90 independence number, 61
Frame quotient, 106 index, 39, 107
(3, 6)-fullerene, 90 inner distribution, 150
fundamental system, 96 integral lattice, 94
integral tree, 85
Gauss sum, 125 interlacing, 31
generalized m-gon, 161 intersection array, 157
generalized n-gon, 114 intersection matrices, 148
generalized adjacency matrix, 175 intersection numbers, 147
generalized line graph, 99
invariant factors, 171
generalized quadrangle, 114
irreducible lattice, 95
generalized quadrangles, 109
irreducible nonnegative matrix, 27
geodetic, 162
isospectral, 22
geometric srg, 109
Geršgorin circles, 36 Johnson graph, 159
Gewirtz graph, 105, 107, 112 join, 31
Godsil’s lemma, 80 just y-cospectral, 175
Google, 55
Gram matrix, 35 k-means, 57
graph, 88 Kelmans operation, 42
graphical Hadamard matrix, 141 Klein graph, 168
Grassmann graph, 159 Kneser graph, 159
Grone-Merris conjecture, 50 Krein conditions, 105, 151
groups (of a design), 107 Krein parameters, 151
Guo conjecture, 49 Kronecker product, 19
Hadamard matrix, 140
half case, 104, 106 Laplace matrix, 11
Hall-Janko graph, 108, 132 Laplace spectrum, 12
halved graph, 114 lattice, 94
Hamiltonian, 69, 70, 112 lattice graph, 18, 109
Hamming graph, 158 level of a rational matrix, 192
Hamming scheme, 155 line graph, 18, 93, 114, 178
Hermitean linear transformation, 27 linear code, 120
Higman-Sims graph, 105 local graph, 188
Higman-Sims technique, 33 loop, 88
Hill code, 124
Hoffman bound, 59, 62 M22 graph, 105, 112
Hoffman coloring, 63 matrix-tree theorem, 15
Hoffman polynomial, 43, 45 Menon design, 141
Hoffman-Singleton graph, 72, 104, metric association scheme, 153
105, 112 minimal polynomial, 166
homomorphism, 88 Moore graph, 113, 153, 161
hypercube, 19 µ-bound, 106, 109
hyperoval, 123 multiplicity free, 88
imprimitive drg, 161 net, 108

228 BIBLIOGRAPHY
Odd graph, 159 Schläfli graph, 65, 110

opposite, 114 Schur complements, 33
orbitals, 106 Schur’s inequality, 33
orthogonal, 115 Seidel adjacency matrix, 23, 142
orthogonal array, 108 Seidel spectrum, 23
orthogonal direct sum, 95 Seidel switching, 23, 142
orthogonality relations, 149 self-dual lattice, 95
semi-bipartite graph, 51
Paley graph, 67, 101, 140, 168 Shannon capacity, 64
partial geometry, 108 Shrikhande graph, 109, 110, 172
partial linear space, 177 SICPOVM, 143
pentagon, 105 sign changes, 154
perfect matching, 72 signless Laplace matrix, 11
period of an irreducible matrix, 28 simplicial complex, 53
Perron-Frobenius, 27 Singleton bound, 123
Perron-Frobenius eigenvector, 55 Smith Normal Form, 171
Petersen graph, 18, 20, 45, 62, 69, spectral center, 84
70, 72, 83, 104, 105, 110, spectral radius, 39
112 spectrum, 11
point graph, 108, 114, 178 split graph, 51
polarity, 77 star complement, 65
positive definite, 35 star partition, 66
positive semidefinite, 35 star subset, 65
primitive nonnegative matrix, 27 stepwise matrix, 42
primitive strongly regular graph, 103 strength, 107
projective code, 120, 122 strictly diagonally dominant, 36
projective plane, 77 strong graph, 133
proper vertex coloring, 62 strongly connected directed graph,
pseudo Latin square graph, 108 28
pseudo-geometric, 109 strongly regular, 101
subdividing, 41
quasisymmetric design, 107
sum graphs, 90
quotient graph, 89
support, 82, 120
quotient matrix, 30
switching class, 23
Ramanujan graph, 67 symmetric design, 76
rank of a permutation group, 106
rationality conditions, 104, 150 t-design, 75
Rayleigh quotient, 31 ternary Golay code, 118, 123
reconstruction conjecture, 24 tesseract, 22
reduced fundamental system, 96 thick generalized n-gon, 114
regular generalized m-gon, 161 threshold graph, 50
regular graph, 13 tight interlacing, 31
regular Hadamard matrix, 141 toughness, 69
regular tournament, 26 tournament, 25
restricted eigenvalue, 103, 133, 197 transitive tournament, 25
root lattice, 93, 96 transversal design, 107
root system, 96 triangular graph, 18, 102, 109, 159
roots, 96 twisted Grassmann graphs, 159
two-graph, 134
Saltire pair, 180, 183, 193 two-weight code, 120
SUBJECT INDEX 229
Type I lattice, 95
Type II lattice, 95
unimodular lattice, 95
unital, 76
universal cover, 88
van Dam-Koolen graphs, 159

vertex connectivity, 110
walk matrix, 192

walk-regular, 195
walks, 14
weak generalized quadrangle, 116
weight, 120
y-cospectral, 175
zero-error channel, 64

Spectra of Graph by Haemers

Uploaded by

Copyright:

Available Formats

Spectra of Graph by Haemers

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Spectra of Graph by Haemers

Uploaded by

Copyright:

Available Formats

Spectra of graphs

Andries E. Brouwer Willem H. Haemers

2.3.1 Equitable and almost equitable partitions of graphs . . . 30

3 Graph spectrum (II) 39

4.5.6 The Petersen graph is not Hamiltonian . . . . . . . . . . 70

6 Groups and graphs 87

8 Strongly regular graphs 101

8.4 Cocliques and colorings . . . . . . . . . . . . . . . . . . . . . . . 112

9 Regular two-graphs 133

10 Association schemes 147

11 Distance regular graphs 157

13 Spectral characterizations 175

14 Graphs with few eigenvalues 195

Algebraic graph theory is the branch of mathematics that studies graphs by

1.1 Matrices associated to a graph

1.2 The spectrum of a graph

(Here, for an eigenvector

Then A has characteristic polynomial pA (θ) = θ3 − 1 and spectrum 1, ω, ω 2 ,

1.2.1 Characteristic polynomial

1.3 The spectrum of an undirected graph

1.3.1 Regular graphs

More generally, we find a lower bound for the diameter:

Proposition 1.3.3 Let Γ be a connected graph with diameter d. Then Γ has at

1.3.5 Spanning trees

Proposition 1.3.4 Let Γ be an undirected (multi)graph with at least one vertex,

2nd proof of Proposition 1.3.4 (sketch) Let Nx be the directed incidence

1.3.6 Bipartite graphs

Proposition 1.3.6 Let Γ be a graph with connected components Γi (1 ≤ i ≤ s).

Proposition 1.3.7 The multiplicity of 0 as a Laplace eigenvalue of an undi-

Proof. We have to show that a connected graph has Laplace eigenvalue 0

Proposition 1.3.8 Let the undirected graph Γ be regular of valency k. Then k

Proposition 1.3.9 The multiplicity of 0 as a signless Laplace eigenvalue of an

Proof. Let M be the vertex-edge incidence matrix of Γ, so that Q = M M ⊤ .

Figure 1.1: Two pairs of cospectral graphs

Proposition 1.3.10 A graph Γ is bipartite if and only if the Laplace spectrum

1.4 Spectrum of some graphs

1.4.1 The complete graph

1.4.2 The complete bipartite graph

1.4.3 The cycle

1.4.4 The path

1.4.5 Line graphs

1.4.6 Cartesian products

If u and v are eigenvectors for Γ and ∆ with ordinary or Laplace eigenvalues θ

1.4.7 Kronecker products and bipartite double

1.4.8 Cayley graphs

1.5.1 Decomposing K10 into Petersen graphs

1.5.2 Decomposing Kn into complete bipartite graphs

Proposition 1.5.1 (H. S. Witsenhausen; Graham & Pollak [162]) Suppose a

Corollary 1.6.1 If all eigenvalues are simple, then Aut Γ is an elementary

Corollary 1.6.2 Let Aut Γ be transitive on X. (Then Γ is regular of degree k,

(i) If m(θ) = 1 for some eigenvalue θ 6= k, then v = |X| is even, and θ ≡ k

(ii) If m(θ) = 1 for two distinct eigenvalues θ 6= k, then v ≡ 0 (mod 4).

1.7 Algebraic connectivity

Proof. Use that µ2 (Γ) = minu {u⊥ Lu | (u, u) = 1, (u, 1) = 0}.

Proposition 1.7.2 Let Γ be a graph with vertex set X. Suppose D ⊂ X is a

Proof. By monotonicity we may assume that Γ contains all edges between

1.8 Cospectral graphs

Figure 1.2: Two cospectral√regular graphs

1.8.1 The 4-cube

1110 1101 1011 1000